]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
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) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com> | |
8 | * | |
9 | * Author: Loic Dachary <loic@dachary.org> | |
10 | * | |
11 | * This is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU Lesser General Public | |
13 | * License version 2.1, as published by the Free Software | |
14 | * Foundation. See file COPYING. | |
15 | * | |
16 | */ | |
17 | ||
224ce89b WB |
18 | #include <boost/algorithm/string.hpp> |
19 | ||
7c673cae FG |
20 | #include "OSDMap.h" |
21 | #include <algorithm> | |
22 | #include "common/config.h" | |
23 | #include "common/Formatter.h" | |
24 | #include "common/TextTable.h" | |
25 | #include "include/ceph_features.h" | |
26 | #include "include/str_map.h" | |
27 | ||
28 | #include "common/code_environment.h" | |
224ce89b | 29 | #include "mon/health_check.h" |
7c673cae FG |
30 | |
31 | #include "crush/CrushTreeDumper.h" | |
32 | #include "common/Clock.h" | |
31f18b77 | 33 | #include "mon/PGStatService.h" |
7c673cae FG |
34 | |
35 | #define dout_subsys ceph_subsys_osd | |
36 | ||
37 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap, osdmap, osdmap); | |
38 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap::Incremental, osdmap_inc, osdmap); | |
39 | ||
40 | ||
41 | // ---------------------------------- | |
42 | // osd_info_t | |
43 | ||
44 | void osd_info_t::dump(Formatter *f) const | |
45 | { | |
46 | f->dump_int("last_clean_begin", last_clean_begin); | |
47 | f->dump_int("last_clean_end", last_clean_end); | |
48 | f->dump_int("up_from", up_from); | |
49 | f->dump_int("up_thru", up_thru); | |
50 | f->dump_int("down_at", down_at); | |
51 | f->dump_int("lost_at", lost_at); | |
52 | } | |
53 | ||
54 | void osd_info_t::encode(bufferlist& bl) const | |
55 | { | |
56 | __u8 struct_v = 1; | |
57 | ::encode(struct_v, bl); | |
58 | ::encode(last_clean_begin, bl); | |
59 | ::encode(last_clean_end, bl); | |
60 | ::encode(up_from, bl); | |
61 | ::encode(up_thru, bl); | |
62 | ::encode(down_at, bl); | |
63 | ::encode(lost_at, bl); | |
64 | } | |
65 | ||
66 | void osd_info_t::decode(bufferlist::iterator& bl) | |
67 | { | |
68 | __u8 struct_v; | |
69 | ::decode(struct_v, bl); | |
70 | ::decode(last_clean_begin, bl); | |
71 | ::decode(last_clean_end, bl); | |
72 | ::decode(up_from, bl); | |
73 | ::decode(up_thru, bl); | |
74 | ::decode(down_at, bl); | |
75 | ::decode(lost_at, bl); | |
76 | } | |
77 | ||
78 | void osd_info_t::generate_test_instances(list<osd_info_t*>& o) | |
79 | { | |
80 | o.push_back(new osd_info_t); | |
81 | o.push_back(new osd_info_t); | |
82 | o.back()->last_clean_begin = 1; | |
83 | o.back()->last_clean_end = 2; | |
84 | o.back()->up_from = 30; | |
85 | o.back()->up_thru = 40; | |
86 | o.back()->down_at = 5; | |
87 | o.back()->lost_at = 6; | |
88 | } | |
89 | ||
90 | ostream& operator<<(ostream& out, const osd_info_t& info) | |
91 | { | |
92 | out << "up_from " << info.up_from | |
93 | << " up_thru " << info.up_thru | |
94 | << " down_at " << info.down_at | |
95 | << " last_clean_interval [" << info.last_clean_begin << "," << info.last_clean_end << ")"; | |
96 | if (info.lost_at) | |
97 | out << " lost_at " << info.lost_at; | |
98 | return out; | |
99 | } | |
100 | ||
101 | // ---------------------------------- | |
102 | // osd_xinfo_t | |
103 | ||
104 | void osd_xinfo_t::dump(Formatter *f) const | |
105 | { | |
106 | f->dump_stream("down_stamp") << down_stamp; | |
107 | f->dump_float("laggy_probability", laggy_probability); | |
108 | f->dump_int("laggy_interval", laggy_interval); | |
109 | f->dump_int("features", features); | |
110 | f->dump_unsigned("old_weight", old_weight); | |
111 | } | |
112 | ||
113 | void osd_xinfo_t::encode(bufferlist& bl) const | |
114 | { | |
115 | ENCODE_START(3, 1, bl); | |
116 | ::encode(down_stamp, bl); | |
117 | __u32 lp = laggy_probability * 0xfffffffful; | |
118 | ::encode(lp, bl); | |
119 | ::encode(laggy_interval, bl); | |
120 | ::encode(features, bl); | |
121 | ::encode(old_weight, bl); | |
122 | ENCODE_FINISH(bl); | |
123 | } | |
124 | ||
125 | void osd_xinfo_t::decode(bufferlist::iterator& bl) | |
126 | { | |
127 | DECODE_START(3, bl); | |
128 | ::decode(down_stamp, bl); | |
129 | __u32 lp; | |
130 | ::decode(lp, bl); | |
131 | laggy_probability = (float)lp / (float)0xffffffff; | |
132 | ::decode(laggy_interval, bl); | |
133 | if (struct_v >= 2) | |
134 | ::decode(features, bl); | |
135 | else | |
136 | features = 0; | |
137 | if (struct_v >= 3) | |
138 | ::decode(old_weight, bl); | |
139 | else | |
140 | old_weight = 0; | |
141 | DECODE_FINISH(bl); | |
142 | } | |
143 | ||
144 | void osd_xinfo_t::generate_test_instances(list<osd_xinfo_t*>& o) | |
145 | { | |
146 | o.push_back(new osd_xinfo_t); | |
147 | o.push_back(new osd_xinfo_t); | |
148 | o.back()->down_stamp = utime_t(2, 3); | |
149 | o.back()->laggy_probability = .123; | |
150 | o.back()->laggy_interval = 123456; | |
151 | o.back()->old_weight = 0x7fff; | |
152 | } | |
153 | ||
154 | ostream& operator<<(ostream& out, const osd_xinfo_t& xi) | |
155 | { | |
156 | return out << "down_stamp " << xi.down_stamp | |
157 | << " laggy_probability " << xi.laggy_probability | |
158 | << " laggy_interval " << xi.laggy_interval | |
159 | << " old_weight " << xi.old_weight; | |
160 | } | |
161 | ||
162 | // ---------------------------------- | |
163 | // OSDMap::Incremental | |
164 | ||
165 | int OSDMap::Incremental::get_net_marked_out(const OSDMap *previous) const | |
166 | { | |
167 | int n = 0; | |
168 | for (auto &weight : new_weight) { | |
169 | if (weight.second == CEPH_OSD_OUT && !previous->is_out(weight.first)) | |
170 | n++; // marked out | |
171 | else if (weight.second != CEPH_OSD_OUT && previous->is_out(weight.first)) | |
172 | n--; // marked in | |
173 | } | |
174 | return n; | |
175 | } | |
176 | ||
177 | int OSDMap::Incremental::get_net_marked_down(const OSDMap *previous) const | |
178 | { | |
179 | int n = 0; | |
180 | for (auto &state : new_state) { // | |
181 | if (state.second & CEPH_OSD_UP) { | |
182 | if (previous->is_up(state.first)) | |
183 | n++; // marked down | |
184 | else | |
185 | n--; // marked up | |
186 | } | |
187 | } | |
188 | return n; | |
189 | } | |
190 | ||
191 | int OSDMap::Incremental::identify_osd(uuid_d u) const | |
192 | { | |
193 | for (auto &uuid : new_uuid) | |
194 | if (uuid.second == u) | |
195 | return uuid.first; | |
196 | return -1; | |
197 | } | |
198 | ||
199 | int OSDMap::Incremental::propagate_snaps_to_tiers(CephContext *cct, | |
200 | const OSDMap& osdmap) | |
201 | { | |
202 | assert(epoch == osdmap.get_epoch() + 1); | |
203 | ||
204 | for (auto &new_pool : new_pools) { | |
205 | if (!new_pool.second.tiers.empty()) { | |
206 | pg_pool_t& base = new_pool.second; | |
207 | ||
208 | for (const auto &tier_pool : base.tiers) { | |
209 | const auto &r = new_pools.find(tier_pool); | |
210 | pg_pool_t *tier = 0; | |
211 | if (r == new_pools.end()) { | |
212 | const pg_pool_t *orig = osdmap.get_pg_pool(tier_pool); | |
213 | if (!orig) { | |
214 | lderr(cct) << __func__ << " no pool " << tier_pool << dendl; | |
215 | return -EIO; | |
216 | } | |
217 | tier = get_new_pool(tier_pool, orig); | |
218 | } else { | |
219 | tier = &r->second; | |
220 | } | |
221 | if (tier->tier_of != new_pool.first) { | |
222 | lderr(cct) << __func__ << " " << r->first << " tier_of != " << new_pool.first << dendl; | |
223 | return -EIO; | |
224 | } | |
225 | ||
226 | ldout(cct, 10) << __func__ << " from " << new_pool.first << " to " | |
227 | << tier_pool << dendl; | |
228 | tier->snap_seq = base.snap_seq; | |
229 | tier->snap_epoch = base.snap_epoch; | |
230 | tier->snaps = base.snaps; | |
231 | tier->removed_snaps = base.removed_snaps; | |
232 | } | |
233 | } | |
234 | } | |
235 | return 0; | |
236 | } | |
237 | ||
238 | ||
239 | bool OSDMap::subtree_is_down(int id, set<int> *down_cache) const | |
240 | { | |
241 | if (id >= 0) | |
242 | return is_down(id); | |
243 | ||
244 | if (down_cache && | |
245 | down_cache->count(id)) { | |
246 | return true; | |
247 | } | |
248 | ||
249 | list<int> children; | |
250 | crush->get_children(id, &children); | |
251 | for (const auto &child : children) { | |
252 | if (!subtree_is_down(child, down_cache)) { | |
253 | return false; | |
254 | } | |
255 | } | |
256 | if (down_cache) { | |
257 | down_cache->insert(id); | |
258 | } | |
259 | return true; | |
260 | } | |
261 | ||
262 | bool OSDMap::containing_subtree_is_down(CephContext *cct, int id, int subtree_type, set<int> *down_cache) const | |
263 | { | |
264 | // use a stack-local down_cache if we didn't get one from the | |
265 | // caller. then at least this particular call will avoid duplicated | |
266 | // work. | |
267 | set<int> local_down_cache; | |
268 | if (!down_cache) { | |
269 | down_cache = &local_down_cache; | |
270 | } | |
271 | ||
272 | int current = id; | |
273 | while (true) { | |
274 | int type; | |
275 | if (current >= 0) { | |
276 | type = 0; | |
277 | } else { | |
278 | type = crush->get_bucket_type(current); | |
279 | } | |
280 | assert(type >= 0); | |
281 | ||
282 | if (!subtree_is_down(current, down_cache)) { | |
283 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = false" << dendl; | |
284 | return false; | |
285 | } | |
286 | ||
287 | // is this a big enough subtree to be marked as down? | |
288 | if (type >= subtree_type) { | |
289 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = true ... " << type << " >= " << subtree_type << dendl; | |
290 | return true; | |
291 | } | |
292 | ||
293 | int r = crush->get_immediate_parent_id(current, ¤t); | |
294 | if (r < 0) { | |
295 | return false; | |
296 | } | |
297 | } | |
298 | } | |
299 | ||
224ce89b WB |
300 | bool OSDMap::subtree_type_is_down( |
301 | CephContext *cct, | |
302 | int id, | |
303 | int subtree_type, | |
304 | set<int> *down_in_osds, | |
305 | set<int> *up_in_osds, | |
306 | set<int> *subtree_up, | |
307 | unordered_map<int, set<int> > *subtree_type_down) const | |
31f18b77 FG |
308 | { |
309 | if (id >= 0) { | |
310 | bool is_down_ret = is_down(id); | |
311 | if (!is_out(id)) { | |
312 | if (is_down_ret) { | |
313 | down_in_osds->insert(id); | |
314 | } else { | |
315 | up_in_osds->insert(id); | |
316 | } | |
317 | } | |
318 | return is_down_ret; | |
319 | } | |
320 | ||
321 | if (subtree_type_down && | |
322 | (*subtree_type_down)[subtree_type].count(id)) { | |
323 | return true; | |
324 | } | |
325 | ||
326 | list<int> children; | |
327 | crush->get_children(id, &children); | |
328 | for (const auto &child : children) { | |
224ce89b WB |
329 | if (!subtree_type_is_down( |
330 | cct, child, crush->get_bucket_type(child), | |
331 | down_in_osds, up_in_osds, subtree_up, subtree_type_down)) { | |
31f18b77 FG |
332 | subtree_up->insert(id); |
333 | return false; | |
334 | } | |
335 | } | |
336 | if (subtree_type_down) { | |
337 | (*subtree_type_down)[subtree_type].insert(id); | |
338 | } | |
339 | return true; | |
340 | } | |
341 | ||
7c673cae FG |
342 | void OSDMap::Incremental::encode_client_old(bufferlist& bl) const |
343 | { | |
344 | __u16 v = 5; | |
345 | ::encode(v, bl); | |
346 | ::encode(fsid, bl); | |
347 | ::encode(epoch, bl); | |
348 | ::encode(modified, bl); | |
349 | int32_t new_t = new_pool_max; | |
350 | ::encode(new_t, bl); | |
351 | ::encode(new_flags, bl); | |
352 | ::encode(fullmap, bl); | |
353 | ::encode(crush, bl); | |
354 | ||
355 | ::encode(new_max_osd, bl); | |
356 | // for ::encode(new_pools, bl); | |
357 | __u32 n = new_pools.size(); | |
358 | ::encode(n, bl); | |
359 | for (const auto &new_pool : new_pools) { | |
360 | n = new_pool.first; | |
361 | ::encode(n, bl); | |
362 | ::encode(new_pool.second, bl, 0); | |
363 | } | |
364 | // for ::encode(new_pool_names, bl); | |
365 | n = new_pool_names.size(); | |
366 | ::encode(n, bl); | |
367 | ||
368 | for (const auto &new_pool_name : new_pool_names) { | |
369 | n = new_pool_name.first; | |
370 | ::encode(n, bl); | |
371 | ::encode(new_pool_name.second, bl); | |
372 | } | |
373 | // for ::encode(old_pools, bl); | |
374 | n = old_pools.size(); | |
375 | ::encode(n, bl); | |
376 | for (auto &old_pool : old_pools) { | |
377 | n = old_pool; | |
378 | ::encode(n, bl); | |
379 | } | |
380 | ::encode(new_up_client, bl, 0); | |
31f18b77 FG |
381 | { |
382 | // legacy is map<int32_t,uint8_t> | |
383 | uint32_t n = new_state.size(); | |
384 | ::encode(n, bl); | |
385 | for (auto p : new_state) { | |
386 | ::encode(p.first, bl); | |
387 | ::encode((uint8_t)p.second, bl); | |
388 | } | |
389 | } | |
7c673cae FG |
390 | ::encode(new_weight, bl); |
391 | // for ::encode(new_pg_temp, bl); | |
392 | n = new_pg_temp.size(); | |
393 | ::encode(n, bl); | |
394 | ||
395 | for (const auto &pg_temp : new_pg_temp) { | |
396 | old_pg_t opg = pg_temp.first.get_old_pg(); | |
397 | ::encode(opg, bl); | |
398 | ::encode(pg_temp.second, bl); | |
399 | } | |
400 | } | |
401 | ||
402 | void OSDMap::Incremental::encode_classic(bufferlist& bl, uint64_t features) const | |
403 | { | |
404 | if ((features & CEPH_FEATURE_PGID64) == 0) { | |
405 | encode_client_old(bl); | |
406 | return; | |
407 | } | |
408 | ||
409 | // base | |
410 | __u16 v = 6; | |
411 | ::encode(v, bl); | |
412 | ::encode(fsid, bl); | |
413 | ::encode(epoch, bl); | |
414 | ::encode(modified, bl); | |
415 | ::encode(new_pool_max, bl); | |
416 | ::encode(new_flags, bl); | |
417 | ::encode(fullmap, bl); | |
418 | ::encode(crush, bl); | |
419 | ||
420 | ::encode(new_max_osd, bl); | |
421 | ::encode(new_pools, bl, features); | |
422 | ::encode(new_pool_names, bl); | |
423 | ::encode(old_pools, bl); | |
424 | ::encode(new_up_client, bl, features); | |
31f18b77 FG |
425 | { |
426 | uint32_t n = new_state.size(); | |
427 | ::encode(n, bl); | |
428 | for (auto p : new_state) { | |
429 | ::encode(p.first, bl); | |
430 | ::encode((uint8_t)p.second, bl); | |
431 | } | |
432 | } | |
7c673cae FG |
433 | ::encode(new_weight, bl); |
434 | ::encode(new_pg_temp, bl); | |
435 | ||
436 | // extended | |
437 | __u16 ev = 10; | |
438 | ::encode(ev, bl); | |
439 | ::encode(new_hb_back_up, bl, features); | |
440 | ::encode(new_up_thru, bl); | |
441 | ::encode(new_last_clean_interval, bl); | |
442 | ::encode(new_lost, bl); | |
443 | ::encode(new_blacklist, bl, features); | |
444 | ::encode(old_blacklist, bl, features); | |
445 | ::encode(new_up_cluster, bl, features); | |
446 | ::encode(cluster_snapshot, bl); | |
447 | ::encode(new_uuid, bl); | |
448 | ::encode(new_xinfo, bl); | |
449 | ::encode(new_hb_front_up, bl, features); | |
450 | } | |
451 | ||
452 | void OSDMap::Incremental::encode(bufferlist& bl, uint64_t features) const | |
453 | { | |
454 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { | |
455 | encode_classic(bl, features); | |
456 | return; | |
457 | } | |
458 | ||
459 | // only a select set of callers should *ever* be encoding new | |
460 | // OSDMaps. others should be passing around the canonical encoded | |
461 | // buffers from on high. select out those callers by passing in an | |
462 | // "impossible" feature bit. | |
463 | assert(features & CEPH_FEATURE_RESERVED); | |
464 | features &= ~CEPH_FEATURE_RESERVED; | |
465 | ||
466 | size_t start_offset = bl.length(); | |
467 | size_t tail_offset; | |
468 | buffer::list::iterator crc_it; | |
469 | ||
470 | // meta-encoding: how we include client-used and osd-specific data | |
471 | ENCODE_START(8, 7, bl); | |
472 | ||
473 | { | |
31f18b77 | 474 | uint8_t v = 5; |
7c673cae FG |
475 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
476 | v = 3; | |
477 | } | |
478 | ENCODE_START(v, 1, bl); // client-usable data | |
479 | ::encode(fsid, bl); | |
480 | ::encode(epoch, bl); | |
481 | ::encode(modified, bl); | |
482 | ::encode(new_pool_max, bl); | |
483 | ::encode(new_flags, bl); | |
484 | ::encode(fullmap, bl); | |
485 | ::encode(crush, bl); | |
486 | ||
487 | ::encode(new_max_osd, bl); | |
488 | ::encode(new_pools, bl, features); | |
489 | ::encode(new_pool_names, bl); | |
490 | ::encode(old_pools, bl); | |
491 | ::encode(new_up_client, bl, features); | |
31f18b77 FG |
492 | if (v >= 5) { |
493 | ::encode(new_state, bl); | |
494 | } else { | |
495 | uint32_t n = new_state.size(); | |
496 | ::encode(n, bl); | |
497 | for (auto p : new_state) { | |
498 | ::encode(p.first, bl); | |
499 | ::encode((uint8_t)p.second, bl); | |
500 | } | |
501 | } | |
7c673cae FG |
502 | ::encode(new_weight, bl); |
503 | ::encode(new_pg_temp, bl); | |
504 | ::encode(new_primary_temp, bl); | |
505 | ::encode(new_primary_affinity, bl); | |
506 | ::encode(new_erasure_code_profiles, bl); | |
507 | ::encode(old_erasure_code_profiles, bl); | |
508 | if (v >= 4) { | |
509 | ::encode(new_pg_upmap, bl); | |
510 | ::encode(old_pg_upmap, bl); | |
511 | ::encode(new_pg_upmap_items, bl); | |
512 | ::encode(old_pg_upmap_items, bl); | |
513 | } | |
514 | ENCODE_FINISH(bl); // client-usable data | |
515 | } | |
516 | ||
517 | { | |
31f18b77 | 518 | uint8_t target_v = 6; |
7c673cae FG |
519 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
520 | target_v = 2; | |
521 | } | |
522 | ENCODE_START(target_v, 1, bl); // extended, osd-only data | |
523 | ::encode(new_hb_back_up, bl, features); | |
524 | ::encode(new_up_thru, bl); | |
525 | ::encode(new_last_clean_interval, bl); | |
526 | ::encode(new_lost, bl); | |
527 | ::encode(new_blacklist, bl, features); | |
528 | ::encode(old_blacklist, bl, features); | |
529 | ::encode(new_up_cluster, bl, features); | |
530 | ::encode(cluster_snapshot, bl); | |
531 | ::encode(new_uuid, bl); | |
532 | ::encode(new_xinfo, bl); | |
533 | ::encode(new_hb_front_up, bl, features); | |
534 | ::encode(features, bl); // NOTE: features arg, not the member | |
535 | if (target_v >= 3) { | |
536 | ::encode(new_nearfull_ratio, bl); | |
537 | ::encode(new_full_ratio, bl); | |
538 | ::encode(new_backfillfull_ratio, bl); | |
31f18b77 FG |
539 | } |
540 | // 5 was string-based new_require_min_compat_client | |
541 | if (target_v >= 6) { | |
7c673cae | 542 | ::encode(new_require_min_compat_client, bl); |
31f18b77 | 543 | ::encode(new_require_osd_release, bl); |
7c673cae FG |
544 | } |
545 | ENCODE_FINISH(bl); // osd-only data | |
546 | } | |
547 | ||
548 | ::encode((uint32_t)0, bl); // dummy inc_crc | |
549 | crc_it = bl.end(); | |
550 | crc_it.advance(-4); | |
551 | tail_offset = bl.length(); | |
552 | ||
553 | ::encode(full_crc, bl); | |
554 | ||
555 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
556 | ||
557 | // fill in crc | |
558 | bufferlist front; | |
559 | front.substr_of(bl, start_offset, crc_it.get_off() - start_offset); | |
560 | inc_crc = front.crc32c(-1); | |
561 | bufferlist tail; | |
562 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); | |
563 | inc_crc = tail.crc32c(inc_crc); | |
564 | ceph_le32 crc_le; | |
565 | crc_le = inc_crc; | |
566 | crc_it.copy_in(4, (char*)&crc_le); | |
567 | have_crc = true; | |
568 | } | |
569 | ||
570 | void OSDMap::Incremental::decode_classic(bufferlist::iterator &p) | |
571 | { | |
572 | __u32 n, t; | |
573 | // base | |
574 | __u16 v; | |
575 | ::decode(v, p); | |
576 | ::decode(fsid, p); | |
577 | ::decode(epoch, p); | |
578 | ::decode(modified, p); | |
579 | if (v == 4 || v == 5) { | |
580 | ::decode(n, p); | |
581 | new_pool_max = n; | |
582 | } else if (v >= 6) | |
583 | ::decode(new_pool_max, p); | |
584 | ::decode(new_flags, p); | |
585 | ::decode(fullmap, p); | |
586 | ::decode(crush, p); | |
587 | ||
588 | ::decode(new_max_osd, p); | |
589 | if (v < 6) { | |
590 | new_pools.clear(); | |
591 | ::decode(n, p); | |
592 | while (n--) { | |
593 | ::decode(t, p); | |
594 | ::decode(new_pools[t], p); | |
595 | } | |
596 | } else { | |
597 | ::decode(new_pools, p); | |
598 | } | |
599 | if (v == 5) { | |
600 | new_pool_names.clear(); | |
601 | ::decode(n, p); | |
602 | while (n--) { | |
603 | ::decode(t, p); | |
604 | ::decode(new_pool_names[t], p); | |
605 | } | |
606 | } else if (v >= 6) { | |
607 | ::decode(new_pool_names, p); | |
608 | } | |
609 | if (v < 6) { | |
610 | old_pools.clear(); | |
611 | ::decode(n, p); | |
612 | while (n--) { | |
613 | ::decode(t, p); | |
614 | old_pools.insert(t); | |
615 | } | |
616 | } else { | |
617 | ::decode(old_pools, p); | |
618 | } | |
619 | ::decode(new_up_client, p); | |
31f18b77 FG |
620 | { |
621 | map<int32_t,uint8_t> ns; | |
622 | ::decode(ns, p); | |
623 | for (auto q : ns) { | |
624 | new_state[q.first] = q.second; | |
625 | } | |
626 | } | |
7c673cae FG |
627 | ::decode(new_weight, p); |
628 | ||
629 | if (v < 6) { | |
630 | new_pg_temp.clear(); | |
631 | ::decode(n, p); | |
632 | while (n--) { | |
633 | old_pg_t opg; | |
634 | ::decode_raw(opg, p); | |
635 | ::decode(new_pg_temp[pg_t(opg)], p); | |
636 | } | |
637 | } else { | |
638 | ::decode(new_pg_temp, p); | |
639 | } | |
640 | ||
641 | // decode short map, too. | |
642 | if (v == 5 && p.end()) | |
643 | return; | |
644 | ||
645 | // extended | |
646 | __u16 ev = 0; | |
647 | if (v >= 5) | |
648 | ::decode(ev, p); | |
649 | ::decode(new_hb_back_up, p); | |
650 | if (v < 5) | |
651 | ::decode(new_pool_names, p); | |
652 | ::decode(new_up_thru, p); | |
653 | ::decode(new_last_clean_interval, p); | |
654 | ::decode(new_lost, p); | |
655 | ::decode(new_blacklist, p); | |
656 | ::decode(old_blacklist, p); | |
657 | if (ev >= 6) | |
658 | ::decode(new_up_cluster, p); | |
659 | if (ev >= 7) | |
660 | ::decode(cluster_snapshot, p); | |
661 | if (ev >= 8) | |
662 | ::decode(new_uuid, p); | |
663 | if (ev >= 9) | |
664 | ::decode(new_xinfo, p); | |
665 | if (ev >= 10) | |
666 | ::decode(new_hb_front_up, p); | |
667 | } | |
668 | ||
669 | void OSDMap::Incremental::decode(bufferlist::iterator& bl) | |
670 | { | |
671 | /** | |
672 | * Older encodings of the Incremental had a single struct_v which | |
673 | * covered the whole encoding, and was prior to our modern | |
674 | * stuff which includes a compatv and a size. So if we see | |
675 | * a struct_v < 7, we must rewind to the beginning and use our | |
676 | * classic decoder. | |
677 | */ | |
678 | size_t start_offset = bl.get_off(); | |
679 | size_t tail_offset = 0; | |
680 | bufferlist crc_front, crc_tail; | |
681 | ||
682 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
683 | if (struct_v < 7) { | |
684 | int struct_v_size = sizeof(struct_v); | |
685 | bl.advance(-struct_v_size); | |
686 | decode_classic(bl); | |
687 | encode_features = 0; | |
688 | if (struct_v >= 6) | |
689 | encode_features = CEPH_FEATURE_PGID64; | |
690 | else | |
691 | encode_features = 0; | |
692 | return; | |
693 | } | |
694 | { | |
31f18b77 | 695 | DECODE_START(5, bl); // client-usable data |
7c673cae FG |
696 | ::decode(fsid, bl); |
697 | ::decode(epoch, bl); | |
698 | ::decode(modified, bl); | |
699 | ::decode(new_pool_max, bl); | |
700 | ::decode(new_flags, bl); | |
701 | ::decode(fullmap, bl); | |
702 | ::decode(crush, bl); | |
703 | ||
704 | ::decode(new_max_osd, bl); | |
705 | ::decode(new_pools, bl); | |
706 | ::decode(new_pool_names, bl); | |
707 | ::decode(old_pools, bl); | |
708 | ::decode(new_up_client, bl); | |
31f18b77 FG |
709 | if (struct_v >= 5) { |
710 | ::decode(new_state, bl); | |
711 | } else { | |
712 | map<int32_t,uint8_t> ns; | |
713 | ::decode(ns, bl); | |
714 | for (auto q : ns) { | |
715 | new_state[q.first] = q.second; | |
716 | } | |
717 | } | |
7c673cae FG |
718 | ::decode(new_weight, bl); |
719 | ::decode(new_pg_temp, bl); | |
720 | ::decode(new_primary_temp, bl); | |
721 | if (struct_v >= 2) | |
722 | ::decode(new_primary_affinity, bl); | |
723 | else | |
724 | new_primary_affinity.clear(); | |
725 | if (struct_v >= 3) { | |
726 | ::decode(new_erasure_code_profiles, bl); | |
727 | ::decode(old_erasure_code_profiles, bl); | |
728 | } else { | |
729 | new_erasure_code_profiles.clear(); | |
730 | old_erasure_code_profiles.clear(); | |
731 | } | |
732 | if (struct_v >= 4) { | |
733 | ::decode(new_pg_upmap, bl); | |
734 | ::decode(old_pg_upmap, bl); | |
735 | ::decode(new_pg_upmap_items, bl); | |
736 | ::decode(old_pg_upmap_items, bl); | |
737 | } | |
738 | DECODE_FINISH(bl); // client-usable data | |
739 | } | |
740 | ||
741 | { | |
31f18b77 | 742 | DECODE_START(6, bl); // extended, osd-only data |
7c673cae FG |
743 | ::decode(new_hb_back_up, bl); |
744 | ::decode(new_up_thru, bl); | |
745 | ::decode(new_last_clean_interval, bl); | |
746 | ::decode(new_lost, bl); | |
747 | ::decode(new_blacklist, bl); | |
748 | ::decode(old_blacklist, bl); | |
749 | ::decode(new_up_cluster, bl); | |
750 | ::decode(cluster_snapshot, bl); | |
751 | ::decode(new_uuid, bl); | |
752 | ::decode(new_xinfo, bl); | |
753 | ::decode(new_hb_front_up, bl); | |
754 | if (struct_v >= 2) | |
755 | ::decode(encode_features, bl); | |
756 | else | |
757 | encode_features = CEPH_FEATURE_PGID64 | CEPH_FEATURE_OSDMAP_ENC; | |
758 | if (struct_v >= 3) { | |
759 | ::decode(new_nearfull_ratio, bl); | |
760 | ::decode(new_full_ratio, bl); | |
761 | } else { | |
762 | new_nearfull_ratio = -1; | |
763 | new_full_ratio = -1; | |
764 | } | |
765 | if (struct_v >= 4) { | |
766 | ::decode(new_backfillfull_ratio, bl); | |
767 | } else { | |
768 | new_backfillfull_ratio = -1; | |
769 | } | |
31f18b77 FG |
770 | if (struct_v == 5) { |
771 | string r; | |
772 | ::decode(r, bl); | |
773 | if (r.length()) { | |
774 | new_require_min_compat_client = ceph_release_from_name(r.c_str()); | |
775 | } | |
776 | } | |
777 | if (struct_v >= 6) { | |
7c673cae | 778 | ::decode(new_require_min_compat_client, bl); |
31f18b77 FG |
779 | ::decode(new_require_osd_release, bl); |
780 | } else { | |
781 | if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_LUMINOUS)) { | |
782 | // only for compat with post-kraken pre-luminous test clusters | |
783 | new_require_osd_release = CEPH_RELEASE_LUMINOUS; | |
784 | new_flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
785 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_KRAKEN)) { | |
786 | new_require_osd_release = CEPH_RELEASE_KRAKEN; | |
787 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_JEWEL)) { | |
788 | new_require_osd_release = CEPH_RELEASE_JEWEL; | |
789 | } else { | |
790 | new_require_osd_release = -1; | |
791 | } | |
792 | } | |
7c673cae FG |
793 | DECODE_FINISH(bl); // osd-only data |
794 | } | |
795 | ||
796 | if (struct_v >= 8) { | |
797 | have_crc = true; | |
798 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
799 | ::decode(inc_crc, bl); | |
800 | tail_offset = bl.get_off(); | |
801 | ::decode(full_crc, bl); | |
802 | } else { | |
803 | have_crc = false; | |
804 | full_crc = 0; | |
805 | inc_crc = 0; | |
806 | } | |
807 | ||
808 | DECODE_FINISH(bl); // wrapper | |
809 | ||
810 | if (have_crc) { | |
811 | // verify crc | |
812 | uint32_t actual = crc_front.crc32c(-1); | |
813 | if (tail_offset < bl.get_off()) { | |
814 | bufferlist tail; | |
815 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); | |
816 | actual = tail.crc32c(actual); | |
817 | } | |
818 | if (inc_crc != actual) { | |
819 | ostringstream ss; | |
820 | ss << "bad crc, actual " << actual << " != expected " << inc_crc; | |
821 | string s = ss.str(); | |
822 | throw buffer::malformed_input(s.c_str()); | |
823 | } | |
824 | } | |
825 | } | |
826 | ||
827 | void OSDMap::Incremental::dump(Formatter *f) const | |
828 | { | |
829 | f->dump_int("epoch", epoch); | |
830 | f->dump_stream("fsid") << fsid; | |
831 | f->dump_stream("modified") << modified; | |
832 | f->dump_int("new_pool_max", new_pool_max); | |
833 | f->dump_int("new_flags", new_flags); | |
834 | f->dump_float("new_full_ratio", new_full_ratio); | |
835 | f->dump_float("new_nearfull_ratio", new_nearfull_ratio); | |
836 | f->dump_float("new_backfillfull_ratio", new_backfillfull_ratio); | |
31f18b77 FG |
837 | f->dump_int("new_require_min_compat_client", new_require_min_compat_client); |
838 | f->dump_int("new_require_osd_release", new_require_osd_release); | |
7c673cae FG |
839 | |
840 | if (fullmap.length()) { | |
841 | f->open_object_section("full_map"); | |
842 | OSDMap full; | |
843 | bufferlist fbl = fullmap; // kludge around constness. | |
844 | auto p = fbl.begin(); | |
845 | full.decode(p); | |
846 | full.dump(f); | |
847 | f->close_section(); | |
848 | } | |
849 | if (crush.length()) { | |
850 | f->open_object_section("crush"); | |
851 | CrushWrapper c; | |
852 | bufferlist tbl = crush; // kludge around constness. | |
853 | auto p = tbl.begin(); | |
854 | c.decode(p); | |
855 | c.dump(f); | |
856 | f->close_section(); | |
857 | } | |
858 | ||
859 | f->dump_int("new_max_osd", new_max_osd); | |
860 | ||
861 | f->open_array_section("new_pools"); | |
862 | ||
863 | for (const auto &new_pool : new_pools) { | |
864 | f->open_object_section("pool"); | |
865 | f->dump_int("pool", new_pool.first); | |
866 | new_pool.second.dump(f); | |
867 | f->close_section(); | |
868 | } | |
869 | f->close_section(); | |
870 | f->open_array_section("new_pool_names"); | |
871 | ||
872 | for (const auto &new_pool_name : new_pool_names) { | |
873 | f->open_object_section("pool_name"); | |
874 | f->dump_int("pool", new_pool_name.first); | |
875 | f->dump_string("name", new_pool_name.second); | |
876 | f->close_section(); | |
877 | } | |
878 | f->close_section(); | |
879 | f->open_array_section("old_pools"); | |
880 | ||
881 | for (const auto &old_pool : old_pools) | |
882 | f->dump_int("pool", old_pool); | |
883 | f->close_section(); | |
884 | ||
885 | f->open_array_section("new_up_osds"); | |
886 | ||
887 | for (const auto &upclient : new_up_client) { | |
888 | f->open_object_section("osd"); | |
889 | f->dump_int("osd", upclient.first); | |
890 | f->dump_stream("public_addr") << upclient.second; | |
891 | f->dump_stream("cluster_addr") << new_up_cluster.find(upclient.first)->second; | |
892 | f->dump_stream("heartbeat_back_addr") << new_hb_back_up.find(upclient.first)->second; | |
893 | map<int32_t, entity_addr_t>::const_iterator q; | |
894 | if ((q = new_hb_front_up.find(upclient.first)) != new_hb_front_up.end()) | |
895 | f->dump_stream("heartbeat_front_addr") << q->second; | |
896 | f->close_section(); | |
897 | } | |
898 | f->close_section(); | |
899 | ||
900 | f->open_array_section("new_weight"); | |
901 | ||
902 | for (const auto &weight : new_weight) { | |
903 | f->open_object_section("osd"); | |
904 | f->dump_int("osd", weight.first); | |
905 | f->dump_int("weight", weight.second); | |
906 | f->close_section(); | |
907 | } | |
908 | f->close_section(); | |
909 | ||
910 | f->open_array_section("osd_state_xor"); | |
911 | for (const auto &ns : new_state) { | |
912 | f->open_object_section("osd"); | |
913 | f->dump_int("osd", ns.first); | |
914 | set<string> st; | |
915 | calc_state_set(new_state.find(ns.first)->second, st); | |
916 | f->open_array_section("state_xor"); | |
917 | for (auto &state : st) | |
918 | f->dump_string("state", state); | |
919 | f->close_section(); | |
c07f9fc5 | 920 | f->close_section(); |
7c673cae FG |
921 | } |
922 | f->close_section(); | |
923 | ||
924 | f->open_array_section("new_pg_temp"); | |
925 | ||
926 | for (const auto &pg_temp : new_pg_temp) { | |
927 | f->open_object_section("pg"); | |
928 | f->dump_stream("pgid") << pg_temp.first; | |
929 | f->open_array_section("osds"); | |
930 | ||
931 | for (const auto &osd : pg_temp.second) | |
932 | f->dump_int("osd", osd); | |
933 | f->close_section(); | |
934 | f->close_section(); | |
935 | } | |
936 | f->close_section(); | |
937 | ||
938 | f->open_array_section("primary_temp"); | |
939 | ||
940 | for (const auto &primary_temp : new_primary_temp) { | |
941 | f->dump_stream("pgid") << primary_temp.first; | |
942 | f->dump_int("osd", primary_temp.second); | |
943 | } | |
944 | f->close_section(); // primary_temp | |
945 | ||
946 | f->open_array_section("new_pg_upmap"); | |
947 | for (auto& i : new_pg_upmap) { | |
948 | f->open_object_section("mapping"); | |
949 | f->dump_stream("pgid") << i.first; | |
950 | f->open_array_section("osds"); | |
951 | for (auto osd : i.second) { | |
952 | f->dump_int("osd", osd); | |
953 | } | |
954 | f->close_section(); | |
955 | f->close_section(); | |
956 | } | |
957 | f->close_section(); | |
958 | f->open_array_section("old_pg_upmap"); | |
959 | for (auto& i : old_pg_upmap) { | |
960 | f->dump_stream("pgid") << i; | |
961 | } | |
962 | f->close_section(); | |
963 | ||
964 | f->open_array_section("new_pg_upmap_items"); | |
965 | for (auto& i : new_pg_upmap_items) { | |
966 | f->open_object_section("mapping"); | |
967 | f->dump_stream("pgid") << i.first; | |
968 | f->open_array_section("mappings"); | |
969 | for (auto& p : i.second) { | |
970 | f->open_object_section("mapping"); | |
971 | f->dump_int("from", p.first); | |
972 | f->dump_int("to", p.second); | |
973 | f->close_section(); | |
974 | } | |
975 | f->close_section(); | |
976 | f->close_section(); | |
977 | } | |
978 | f->close_section(); | |
979 | f->open_array_section("old_pg_upmap_items"); | |
980 | for (auto& i : old_pg_upmap_items) { | |
981 | f->dump_stream("pgid") << i; | |
982 | } | |
983 | f->close_section(); | |
984 | ||
985 | f->open_array_section("new_up_thru"); | |
986 | ||
987 | for (const auto &up_thru : new_up_thru) { | |
988 | f->open_object_section("osd"); | |
989 | f->dump_int("osd", up_thru.first); | |
990 | f->dump_int("up_thru", up_thru.second); | |
991 | f->close_section(); | |
992 | } | |
993 | f->close_section(); | |
994 | ||
995 | f->open_array_section("new_lost"); | |
996 | ||
997 | for (const auto &lost : new_lost) { | |
998 | f->open_object_section("osd"); | |
999 | f->dump_int("osd", lost.first); | |
1000 | f->dump_int("epoch_lost", lost.second); | |
1001 | f->close_section(); | |
1002 | } | |
1003 | f->close_section(); | |
1004 | ||
1005 | f->open_array_section("new_last_clean_interval"); | |
1006 | ||
1007 | for (const auto &last_clean_interval : new_last_clean_interval) { | |
1008 | f->open_object_section("osd"); | |
1009 | f->dump_int("osd", last_clean_interval.first); | |
1010 | f->dump_int("first", last_clean_interval.second.first); | |
1011 | f->dump_int("last", last_clean_interval.second.second); | |
1012 | f->close_section(); | |
1013 | } | |
1014 | f->close_section(); | |
1015 | ||
1016 | f->open_array_section("new_blacklist"); | |
1017 | for (const auto &blist : new_blacklist) { | |
1018 | stringstream ss; | |
1019 | ss << blist.first; | |
1020 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1021 | } | |
1022 | f->close_section(); | |
1023 | f->open_array_section("old_blacklist"); | |
1024 | for (const auto &blist : old_blacklist) | |
1025 | f->dump_stream("addr") << blist; | |
1026 | f->close_section(); | |
1027 | ||
1028 | f->open_array_section("new_xinfo"); | |
1029 | for (const auto &xinfo : new_xinfo) { | |
1030 | f->open_object_section("xinfo"); | |
1031 | f->dump_int("osd", xinfo.first); | |
1032 | xinfo.second.dump(f); | |
1033 | f->close_section(); | |
1034 | } | |
1035 | f->close_section(); | |
1036 | ||
1037 | if (cluster_snapshot.size()) | |
1038 | f->dump_string("cluster_snapshot", cluster_snapshot); | |
1039 | ||
1040 | f->open_array_section("new_uuid"); | |
1041 | for (const auto &uuid : new_uuid) { | |
1042 | f->open_object_section("osd"); | |
1043 | f->dump_int("osd", uuid.first); | |
1044 | f->dump_stream("uuid") << uuid.second; | |
1045 | f->close_section(); | |
1046 | } | |
1047 | f->close_section(); | |
1048 | ||
1049 | OSDMap::dump_erasure_code_profiles(new_erasure_code_profiles, f); | |
1050 | f->open_array_section("old_erasure_code_profiles"); | |
1051 | for (const auto &erasure_code_profile : old_erasure_code_profiles) { | |
1052 | f->dump_string("old", erasure_code_profile.c_str()); | |
1053 | } | |
1054 | f->close_section(); | |
1055 | } | |
1056 | ||
1057 | void OSDMap::Incremental::generate_test_instances(list<Incremental*>& o) | |
1058 | { | |
1059 | o.push_back(new Incremental); | |
1060 | } | |
1061 | ||
1062 | // ---------------------------------- | |
1063 | // OSDMap | |
1064 | ||
1065 | void OSDMap::set_epoch(epoch_t e) | |
1066 | { | |
1067 | epoch = e; | |
1068 | for (auto &pool : pools) | |
1069 | pool.second.last_change = e; | |
1070 | } | |
1071 | ||
1072 | bool OSDMap::is_blacklisted(const entity_addr_t& a) const | |
1073 | { | |
1074 | if (blacklist.empty()) | |
1075 | return false; | |
1076 | ||
1077 | // this specific instance? | |
1078 | if (blacklist.count(a)) | |
1079 | return true; | |
1080 | ||
1081 | // is entire ip blacklisted? | |
1082 | if (a.is_ip()) { | |
1083 | entity_addr_t b = a; | |
1084 | b.set_port(0); | |
1085 | b.set_nonce(0); | |
1086 | if (blacklist.count(b)) { | |
1087 | return true; | |
1088 | } | |
1089 | } | |
1090 | ||
1091 | return false; | |
1092 | } | |
1093 | ||
1094 | void OSDMap::get_blacklist(list<pair<entity_addr_t,utime_t> > *bl) const | |
1095 | { | |
1096 | std::copy(blacklist.begin(), blacklist.end(), std::back_inserter(*bl)); | |
1097 | } | |
1098 | ||
31f18b77 FG |
1099 | void OSDMap::get_blacklist(std::set<entity_addr_t> *bl) const |
1100 | { | |
1101 | for (const auto &i : blacklist) { | |
1102 | bl->insert(i.first); | |
1103 | } | |
1104 | } | |
1105 | ||
7c673cae FG |
1106 | void OSDMap::set_max_osd(int m) |
1107 | { | |
1108 | int o = max_osd; | |
1109 | max_osd = m; | |
1110 | osd_state.resize(m); | |
1111 | osd_weight.resize(m); | |
1112 | for (; o<max_osd; o++) { | |
1113 | osd_state[o] = 0; | |
1114 | osd_weight[o] = CEPH_OSD_OUT; | |
1115 | } | |
1116 | osd_info.resize(m); | |
1117 | osd_xinfo.resize(m); | |
1118 | osd_addrs->client_addr.resize(m); | |
1119 | osd_addrs->cluster_addr.resize(m); | |
1120 | osd_addrs->hb_back_addr.resize(m); | |
1121 | osd_addrs->hb_front_addr.resize(m); | |
1122 | osd_uuid->resize(m); | |
1123 | if (osd_primary_affinity) | |
1124 | osd_primary_affinity->resize(m, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
1125 | ||
1126 | calc_num_osds(); | |
1127 | } | |
1128 | ||
1129 | int OSDMap::calc_num_osds() | |
1130 | { | |
1131 | num_osd = 0; | |
1132 | num_up_osd = 0; | |
1133 | num_in_osd = 0; | |
1134 | for (int i=0; i<max_osd; i++) { | |
1135 | if (osd_state[i] & CEPH_OSD_EXISTS) { | |
1136 | ++num_osd; | |
1137 | if (osd_state[i] & CEPH_OSD_UP) { | |
1138 | ++num_up_osd; | |
1139 | } | |
1140 | if (get_weight(i) != CEPH_OSD_OUT) { | |
1141 | ++num_in_osd; | |
1142 | } | |
1143 | } | |
1144 | } | |
1145 | return num_osd; | |
1146 | } | |
1147 | ||
1148 | void OSDMap::count_full_nearfull_osds(int *full, int *backfill, int *nearfull) const | |
1149 | { | |
1150 | *full = 0; | |
1151 | *backfill = 0; | |
1152 | *nearfull = 0; | |
1153 | for (int i = 0; i < max_osd; ++i) { | |
1154 | if (exists(i) && is_up(i) && is_in(i)) { | |
1155 | if (osd_state[i] & CEPH_OSD_FULL) | |
1156 | ++(*full); | |
1157 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1158 | ++(*backfill); | |
1159 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1160 | ++(*nearfull); | |
1161 | } | |
1162 | } | |
1163 | } | |
1164 | ||
31f18b77 FG |
1165 | static bool get_osd_utilization( |
1166 | const mempool::pgmap::unordered_map<int32_t,osd_stat_t> &osd_stat, | |
1167 | int id, int64_t* kb, int64_t* kb_used, int64_t* kb_avail) | |
1168 | { | |
1169 | auto p = osd_stat.find(id); | |
1170 | if (p == osd_stat.end()) | |
1171 | return false; | |
1172 | *kb = p->second.kb; | |
1173 | *kb_used = p->second.kb_used; | |
1174 | *kb_avail = p->second.kb_avail; | |
1175 | return *kb > 0; | |
7c673cae FG |
1176 | } |
1177 | ||
31f18b77 FG |
1178 | void OSDMap::get_full_osd_util( |
1179 | const mempool::pgmap::unordered_map<int32_t,osd_stat_t> &osd_stat, | |
1180 | map<int, float> *full, map<int, float> *backfill, map<int, float> *nearfull) const | |
7c673cae FG |
1181 | { |
1182 | full->clear(); | |
1183 | backfill->clear(); | |
1184 | nearfull->clear(); | |
1185 | for (int i = 0; i < max_osd; ++i) { | |
1186 | if (exists(i) && is_up(i) && is_in(i)) { | |
1187 | int64_t kb, kb_used, kb_avail; | |
1188 | if (osd_state[i] & CEPH_OSD_FULL) { | |
1189 | if (get_osd_utilization(osd_stat, i, &kb, &kb_used, &kb_avail)) | |
1190 | full->emplace(i, (float)kb_used / (float)kb); | |
1191 | } else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) { | |
1192 | if (get_osd_utilization(osd_stat, i, &kb, &kb_used, &kb_avail)) | |
1193 | backfill->emplace(i, (float)kb_used / (float)kb); | |
1194 | } else if (osd_state[i] & CEPH_OSD_NEARFULL) { | |
1195 | if (get_osd_utilization(osd_stat, i, &kb, &kb_used, &kb_avail)) | |
1196 | nearfull->emplace(i, (float)kb_used / (float)kb); | |
1197 | } | |
1198 | } | |
1199 | } | |
1200 | } | |
1201 | ||
31f18b77 FG |
1202 | void OSDMap::get_full_osd_counts(set<int> *full, set<int> *backfill, |
1203 | set<int> *nearfull) const | |
1204 | { | |
1205 | full->clear(); | |
1206 | backfill->clear(); | |
1207 | nearfull->clear(); | |
1208 | for (int i = 0; i < max_osd; ++i) { | |
1209 | if (exists(i) && is_up(i) && is_in(i)) { | |
1210 | if (osd_state[i] & CEPH_OSD_FULL) | |
1211 | full->emplace(i); | |
1212 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1213 | backfill->emplace(i); | |
1214 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1215 | nearfull->emplace(i); | |
1216 | } | |
1217 | } | |
1218 | } | |
1219 | ||
7c673cae FG |
1220 | void OSDMap::get_all_osds(set<int32_t>& ls) const |
1221 | { | |
1222 | for (int i=0; i<max_osd; i++) | |
1223 | if (exists(i)) | |
1224 | ls.insert(i); | |
1225 | } | |
1226 | ||
1227 | void OSDMap::get_up_osds(set<int32_t>& ls) const | |
1228 | { | |
1229 | for (int i = 0; i < max_osd; i++) { | |
1230 | if (is_up(i)) | |
1231 | ls.insert(i); | |
1232 | } | |
1233 | } | |
1234 | ||
31f18b77 FG |
1235 | void OSDMap::get_out_osds(set<int32_t>& ls) const |
1236 | { | |
1237 | for (int i = 0; i < max_osd; i++) { | |
1238 | if (is_out(i)) | |
1239 | ls.insert(i); | |
1240 | } | |
1241 | } | |
1242 | ||
7c673cae FG |
1243 | void OSDMap::calc_state_set(int state, set<string>& st) |
1244 | { | |
1245 | unsigned t = state; | |
1246 | for (unsigned s = 1; t; s <<= 1) { | |
1247 | if (t & s) { | |
1248 | t &= ~s; | |
1249 | st.insert(ceph_osd_state_name(s)); | |
1250 | } | |
1251 | } | |
1252 | } | |
1253 | ||
1254 | void OSDMap::adjust_osd_weights(const map<int,double>& weights, Incremental& inc) const | |
1255 | { | |
1256 | float max = 0; | |
1257 | for (const auto &weight : weights) { | |
1258 | if (weight.second > max) | |
1259 | max = weight.second; | |
1260 | } | |
1261 | ||
1262 | for (const auto &weight : weights) { | |
1263 | inc.new_weight[weight.first] = (unsigned)((weight.second / max) * CEPH_OSD_IN); | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | int OSDMap::identify_osd(const entity_addr_t& addr) const | |
1268 | { | |
1269 | for (int i=0; i<max_osd; i++) | |
1270 | if (exists(i) && (get_addr(i) == addr || get_cluster_addr(i) == addr)) | |
1271 | return i; | |
1272 | return -1; | |
1273 | } | |
1274 | ||
1275 | int OSDMap::identify_osd(const uuid_d& u) const | |
1276 | { | |
1277 | for (int i=0; i<max_osd; i++) | |
1278 | if (exists(i) && get_uuid(i) == u) | |
1279 | return i; | |
1280 | return -1; | |
1281 | } | |
1282 | ||
1283 | int OSDMap::identify_osd_on_all_channels(const entity_addr_t& addr) const | |
1284 | { | |
1285 | for (int i=0; i<max_osd; i++) | |
1286 | if (exists(i) && (get_addr(i) == addr || get_cluster_addr(i) == addr || | |
1287 | get_hb_back_addr(i) == addr || get_hb_front_addr(i) == addr)) | |
1288 | return i; | |
1289 | return -1; | |
1290 | } | |
1291 | ||
1292 | int OSDMap::find_osd_on_ip(const entity_addr_t& ip) const | |
1293 | { | |
1294 | for (int i=0; i<max_osd; i++) | |
1295 | if (exists(i) && (get_addr(i).is_same_host(ip) || get_cluster_addr(i).is_same_host(ip))) | |
1296 | return i; | |
1297 | return -1; | |
1298 | } | |
1299 | ||
1300 | ||
1301 | uint64_t OSDMap::get_features(int entity_type, uint64_t *pmask) const | |
1302 | { | |
1303 | uint64_t features = 0; // things we actually have | |
1304 | uint64_t mask = 0; // things we could have | |
1305 | ||
1306 | if (crush->has_nondefault_tunables()) | |
1307 | features |= CEPH_FEATURE_CRUSH_TUNABLES; | |
1308 | if (crush->has_nondefault_tunables2()) | |
1309 | features |= CEPH_FEATURE_CRUSH_TUNABLES2; | |
1310 | if (crush->has_nondefault_tunables3()) | |
1311 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1312 | if (crush->has_v4_buckets()) | |
1313 | features |= CEPH_FEATURE_CRUSH_V4; | |
1314 | if (crush->has_nondefault_tunables5()) | |
1315 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
c07f9fc5 FG |
1316 | if (crush->has_incompat_choose_args()) { |
1317 | features |= CEPH_FEATUREMASK_CRUSH_CHOOSE_ARGS; | |
1318 | } | |
7c673cae FG |
1319 | mask |= CEPH_FEATURES_CRUSH; |
1320 | ||
1321 | if (!pg_upmap.empty() || !pg_upmap_items.empty()) | |
1322 | features |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1323 | mask |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1324 | ||
1325 | for (auto &pool: pools) { | |
1326 | if (pool.second.has_flag(pg_pool_t::FLAG_HASHPSPOOL)) { | |
1327 | features |= CEPH_FEATURE_OSDHASHPSPOOL; | |
1328 | } | |
1329 | if (pool.second.is_erasure() && | |
1330 | entity_type != CEPH_ENTITY_TYPE_CLIENT) { // not for clients | |
1331 | features |= CEPH_FEATURE_OSD_ERASURE_CODES; | |
1332 | } | |
1333 | if (!pool.second.tiers.empty() || | |
1334 | pool.second.is_tier()) { | |
1335 | features |= CEPH_FEATURE_OSD_CACHEPOOL; | |
1336 | } | |
31f18b77 | 1337 | int ruleid = crush->find_rule(pool.second.get_crush_rule(), |
7c673cae FG |
1338 | pool.second.get_type(), |
1339 | pool.second.get_size()); | |
1340 | if (ruleid >= 0) { | |
1341 | if (crush->is_v2_rule(ruleid)) | |
1342 | features |= CEPH_FEATURE_CRUSH_V2; | |
1343 | if (crush->is_v3_rule(ruleid)) | |
1344 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1345 | if (crush->is_v5_rule(ruleid)) | |
1346 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
1347 | } | |
1348 | } | |
1349 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1350 | for (auto &erasure_code_profile : erasure_code_profiles) { | |
1351 | auto& profile = erasure_code_profile.second; | |
1352 | const auto& plugin = profile.find("plugin"); | |
1353 | if (plugin != profile.end()) { | |
1354 | if (plugin->second == "isa" || plugin->second == "lrc") | |
1355 | features |= CEPH_FEATURE_ERASURE_CODE_PLUGINS_V2; | |
1356 | if (plugin->second == "shec") | |
1357 | features |= CEPH_FEATURE_ERASURE_CODE_PLUGINS_V3; | |
1358 | } | |
1359 | } | |
1360 | } | |
1361 | mask |= CEPH_FEATURE_OSDHASHPSPOOL | CEPH_FEATURE_OSD_CACHEPOOL; | |
1362 | if (entity_type != CEPH_ENTITY_TYPE_CLIENT) | |
1363 | mask |= CEPH_FEATURE_OSD_ERASURE_CODES; | |
1364 | ||
1365 | if (osd_primary_affinity) { | |
1366 | for (int i = 0; i < max_osd; ++i) { | |
1367 | if ((*osd_primary_affinity)[i] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
1368 | features |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1369 | break; | |
1370 | } | |
1371 | } | |
1372 | } | |
1373 | mask |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1374 | ||
1375 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1376 | const uint64_t jewel_features = CEPH_FEATURE_SERVER_JEWEL; | |
31f18b77 | 1377 | if (require_osd_release >= CEPH_RELEASE_JEWEL) { |
7c673cae FG |
1378 | features |= jewel_features; |
1379 | } | |
1380 | mask |= jewel_features; | |
1381 | ||
1382 | const uint64_t kraken_features = CEPH_FEATUREMASK_SERVER_KRAKEN | |
1383 | | CEPH_FEATURE_MSG_ADDR2; | |
31f18b77 | 1384 | if (require_osd_release >= CEPH_RELEASE_KRAKEN) { |
7c673cae FG |
1385 | features |= kraken_features; |
1386 | } | |
1387 | mask |= kraken_features; | |
1388 | } | |
1389 | ||
1390 | if (pmask) | |
1391 | *pmask = mask; | |
1392 | return features; | |
1393 | } | |
1394 | ||
31f18b77 | 1395 | uint8_t OSDMap::get_min_compat_client() const |
7c673cae FG |
1396 | { |
1397 | uint64_t f = get_features(CEPH_ENTITY_TYPE_CLIENT, nullptr); | |
1398 | ||
1399 | if (HAVE_FEATURE(f, OSDMAP_PG_UPMAP) || // v12.0.0-1733-g27d6f43 | |
31f18b77 FG |
1400 | HAVE_FEATURE(f, CRUSH_CHOOSE_ARGS)) { // v12.0.1-2172-gef1ef28 |
1401 | return CEPH_RELEASE_LUMINOUS; // v12.2.0 | |
7c673cae FG |
1402 | } |
1403 | if (HAVE_FEATURE(f, CRUSH_TUNABLES5)) { // v10.0.0-612-g043a737 | |
31f18b77 | 1404 | return CEPH_RELEASE_JEWEL; // v10.2.0 |
7c673cae FG |
1405 | } |
1406 | if (HAVE_FEATURE(f, CRUSH_V4)) { // v0.91-678-g325fc56 | |
31f18b77 | 1407 | return CEPH_RELEASE_HAMMER; // v0.94.0 |
7c673cae FG |
1408 | } |
1409 | if (HAVE_FEATURE(f, OSD_PRIMARY_AFFINITY) || // v0.76-553-gf825624 | |
1410 | HAVE_FEATURE(f, CRUSH_TUNABLES3) || // v0.76-395-ge20a55d | |
1411 | HAVE_FEATURE(f, OSD_ERASURE_CODES) || // v0.73-498-gbfc86a8 | |
1412 | HAVE_FEATURE(f, OSD_CACHEPOOL)) { // v0.67-401-gb91c1c5 | |
31f18b77 | 1413 | return CEPH_RELEASE_FIREFLY; // v0.80.0 |
7c673cae FG |
1414 | } |
1415 | if (HAVE_FEATURE(f, CRUSH_TUNABLES2) || // v0.54-684-g0cc47ff | |
1416 | HAVE_FEATURE(f, OSDHASHPSPOOL)) { // v0.57-398-g8cc2b0f | |
31f18b77 | 1417 | return CEPH_RELEASE_DUMPLING; // v0.67.0 |
7c673cae FG |
1418 | } |
1419 | if (HAVE_FEATURE(f, CRUSH_TUNABLES)) { // v0.48argonaut-206-g6f381af | |
31f18b77 | 1420 | return CEPH_RELEASE_ARGONAUT; // v0.48argonaut-206-g6f381af |
7c673cae | 1421 | } |
31f18b77 | 1422 | return CEPH_RELEASE_ARGONAUT; // v0.48argonaut-206-g6f381af |
7c673cae FG |
1423 | } |
1424 | ||
1425 | void OSDMap::_calc_up_osd_features() | |
1426 | { | |
1427 | bool first = true; | |
1428 | cached_up_osd_features = 0; | |
1429 | for (int osd = 0; osd < max_osd; ++osd) { | |
1430 | if (!is_up(osd)) | |
1431 | continue; | |
1432 | const osd_xinfo_t &xi = get_xinfo(osd); | |
1433 | if (first) { | |
1434 | cached_up_osd_features = xi.features; | |
1435 | first = false; | |
1436 | } else { | |
1437 | cached_up_osd_features &= xi.features; | |
1438 | } | |
1439 | } | |
1440 | } | |
1441 | ||
1442 | uint64_t OSDMap::get_up_osd_features() const | |
1443 | { | |
1444 | return cached_up_osd_features; | |
1445 | } | |
1446 | ||
1447 | void OSDMap::dedup(const OSDMap *o, OSDMap *n) | |
1448 | { | |
1449 | if (o->epoch == n->epoch) | |
1450 | return; | |
1451 | ||
1452 | int diff = 0; | |
1453 | ||
1454 | // do addrs match? | |
1455 | if (o->max_osd != n->max_osd) | |
1456 | diff++; | |
1457 | for (int i = 0; i < o->max_osd && i < n->max_osd; i++) { | |
1458 | if ( n->osd_addrs->client_addr[i] && o->osd_addrs->client_addr[i] && | |
1459 | *n->osd_addrs->client_addr[i] == *o->osd_addrs->client_addr[i]) | |
1460 | n->osd_addrs->client_addr[i] = o->osd_addrs->client_addr[i]; | |
1461 | else | |
1462 | diff++; | |
1463 | if ( n->osd_addrs->cluster_addr[i] && o->osd_addrs->cluster_addr[i] && | |
1464 | *n->osd_addrs->cluster_addr[i] == *o->osd_addrs->cluster_addr[i]) | |
1465 | n->osd_addrs->cluster_addr[i] = o->osd_addrs->cluster_addr[i]; | |
1466 | else | |
1467 | diff++; | |
1468 | if ( n->osd_addrs->hb_back_addr[i] && o->osd_addrs->hb_back_addr[i] && | |
1469 | *n->osd_addrs->hb_back_addr[i] == *o->osd_addrs->hb_back_addr[i]) | |
1470 | n->osd_addrs->hb_back_addr[i] = o->osd_addrs->hb_back_addr[i]; | |
1471 | else | |
1472 | diff++; | |
1473 | if ( n->osd_addrs->hb_front_addr[i] && o->osd_addrs->hb_front_addr[i] && | |
1474 | *n->osd_addrs->hb_front_addr[i] == *o->osd_addrs->hb_front_addr[i]) | |
1475 | n->osd_addrs->hb_front_addr[i] = o->osd_addrs->hb_front_addr[i]; | |
1476 | else | |
1477 | diff++; | |
1478 | } | |
1479 | if (diff == 0) { | |
1480 | // zoinks, no differences at all! | |
1481 | n->osd_addrs = o->osd_addrs; | |
1482 | } | |
1483 | ||
1484 | // does crush match? | |
1485 | bufferlist oc, nc; | |
1486 | ::encode(*o->crush, oc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
1487 | ::encode(*n->crush, nc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
1488 | if (oc.contents_equal(nc)) { | |
1489 | n->crush = o->crush; | |
1490 | } | |
1491 | ||
1492 | // does pg_temp match? | |
31f18b77 FG |
1493 | if (*o->pg_temp == *n->pg_temp) |
1494 | n->pg_temp = o->pg_temp; | |
7c673cae FG |
1495 | |
1496 | // does primary_temp match? | |
1497 | if (o->primary_temp->size() == n->primary_temp->size()) { | |
1498 | if (*o->primary_temp == *n->primary_temp) | |
1499 | n->primary_temp = o->primary_temp; | |
1500 | } | |
1501 | ||
1502 | // do uuids match? | |
1503 | if (o->osd_uuid->size() == n->osd_uuid->size() && | |
1504 | *o->osd_uuid == *n->osd_uuid) | |
1505 | n->osd_uuid = o->osd_uuid; | |
1506 | } | |
1507 | ||
1508 | void OSDMap::clean_temps(CephContext *cct, | |
1509 | const OSDMap& osdmap, Incremental *pending_inc) | |
1510 | { | |
1511 | ldout(cct, 10) << __func__ << dendl; | |
1512 | OSDMap tmpmap; | |
1513 | tmpmap.deepish_copy_from(osdmap); | |
1514 | tmpmap.apply_incremental(*pending_inc); | |
1515 | ||
1516 | for (auto pg : *tmpmap.pg_temp) { | |
1517 | // if pool does not exist, remove any existing pg_temps associated with | |
1518 | // it. we don't care about pg_temps on the pending_inc either; if there | |
1519 | // are new_pg_temp entries on the pending, clear them out just as well. | |
1520 | if (!osdmap.have_pg_pool(pg.first.pool())) { | |
1521 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1522 | << " for nonexistent pool " << pg.first.pool() << dendl; | |
1523 | pending_inc->new_pg_temp[pg.first].clear(); | |
1524 | continue; | |
1525 | } | |
1526 | // all osds down? | |
1527 | unsigned num_up = 0; | |
1528 | for (auto o : pg.second) { | |
1529 | if (!tmpmap.is_down(o)) { | |
1530 | ++num_up; | |
1531 | break; | |
1532 | } | |
1533 | } | |
1534 | if (num_up == 0) { | |
1535 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1536 | << " with all down osds" << pg.second << dendl; | |
1537 | pending_inc->new_pg_temp[pg.first].clear(); | |
1538 | continue; | |
1539 | } | |
1540 | // redundant pg_temp? | |
1541 | vector<int> raw_up; | |
1542 | int primary; | |
1543 | tmpmap.pg_to_raw_up(pg.first, &raw_up, &primary); | |
1544 | if (vectors_equal(raw_up, pg.second)) { | |
1545 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " | |
1546 | << pg.second << " that matches raw_up mapping" << dendl; | |
1547 | if (osdmap.pg_temp->count(pg.first)) | |
1548 | pending_inc->new_pg_temp[pg.first].clear(); | |
1549 | else | |
1550 | pending_inc->new_pg_temp.erase(pg.first); | |
1551 | } | |
1552 | } | |
1553 | ||
1554 | for (auto &pg : *tmpmap.primary_temp) { | |
1555 | // primary down? | |
1556 | if (tmpmap.is_down(pg.second)) { | |
1557 | ldout(cct, 10) << __func__ << " removing primary_temp " << pg.first | |
1558 | << " to down " << pg.second << dendl; | |
1559 | pending_inc->new_primary_temp[pg.first] = -1; | |
1560 | continue; | |
1561 | } | |
1562 | // redundant primary_temp? | |
1563 | vector<int> real_up, templess_up; | |
1564 | int real_primary, templess_primary; | |
1565 | pg_t pgid = pg.first; | |
1566 | tmpmap.pg_to_acting_osds(pgid, &real_up, &real_primary); | |
1567 | tmpmap.pg_to_raw_up(pgid, &templess_up, &templess_primary); | |
1568 | if (real_primary == templess_primary){ | |
1569 | ldout(cct, 10) << __func__ << " removing primary_temp " | |
1570 | << pgid << " -> " << real_primary | |
1571 | << " (unnecessary/redundant)" << dendl; | |
1572 | if (osdmap.primary_temp->count(pgid)) | |
1573 | pending_inc->new_primary_temp[pgid] = -1; | |
1574 | else | |
1575 | pending_inc->new_primary_temp.erase(pgid); | |
1576 | } | |
1577 | } | |
1578 | } | |
1579 | ||
1580 | int OSDMap::apply_incremental(const Incremental &inc) | |
1581 | { | |
1582 | new_blacklist_entries = false; | |
1583 | if (inc.epoch == 1) | |
1584 | fsid = inc.fsid; | |
1585 | else if (inc.fsid != fsid) | |
1586 | return -EINVAL; | |
1587 | ||
1588 | assert(inc.epoch == epoch+1); | |
1589 | ||
1590 | epoch++; | |
1591 | modified = inc.modified; | |
1592 | ||
1593 | // full map? | |
1594 | if (inc.fullmap.length()) { | |
1595 | bufferlist bl(inc.fullmap); | |
1596 | decode(bl); | |
1597 | return 0; | |
1598 | } | |
1599 | ||
1600 | // nope, incremental. | |
31f18b77 | 1601 | if (inc.new_flags >= 0) { |
7c673cae | 1602 | flags = inc.new_flags; |
31f18b77 FG |
1603 | // the below is just to cover a newly-upgraded luminous mon |
1604 | // cluster that has to set require_jewel_osds or | |
1605 | // require_kraken_osds before the osds can be upgraded to | |
1606 | // luminous. | |
1607 | if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { | |
1608 | if (require_osd_release < CEPH_RELEASE_KRAKEN) { | |
1609 | require_osd_release = CEPH_RELEASE_KRAKEN; | |
1610 | } | |
1611 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
1612 | if (require_osd_release < CEPH_RELEASE_JEWEL) { | |
1613 | require_osd_release = CEPH_RELEASE_JEWEL; | |
1614 | } | |
1615 | } | |
1616 | } | |
7c673cae FG |
1617 | |
1618 | if (inc.new_max_osd >= 0) | |
1619 | set_max_osd(inc.new_max_osd); | |
1620 | ||
1621 | if (inc.new_pool_max != -1) | |
1622 | pool_max = inc.new_pool_max; | |
1623 | ||
1624 | for (const auto &pool : inc.new_pools) { | |
1625 | pools[pool.first] = pool.second; | |
1626 | pools[pool.first].last_change = epoch; | |
1627 | } | |
1628 | ||
1629 | for (const auto &pname : inc.new_pool_names) { | |
1630 | auto pool_name_entry = pool_name.find(pname.first); | |
1631 | if (pool_name_entry != pool_name.end()) { | |
1632 | name_pool.erase(pool_name_entry->second); | |
1633 | pool_name_entry->second = pname.second; | |
1634 | } else { | |
1635 | pool_name[pname.first] = pname.second; | |
1636 | } | |
1637 | name_pool[pname.second] = pname.first; | |
1638 | } | |
1639 | ||
1640 | for (const auto &pool : inc.old_pools) { | |
1641 | pools.erase(pool); | |
1642 | name_pool.erase(pool_name[pool]); | |
1643 | pool_name.erase(pool); | |
1644 | } | |
1645 | ||
1646 | for (const auto &weight : inc.new_weight) { | |
1647 | set_weight(weight.first, weight.second); | |
1648 | ||
1649 | // if we are marking in, clear the AUTOOUT and NEW bits, and clear | |
1650 | // xinfo old_weight. | |
1651 | if (weight.second) { | |
1652 | osd_state[weight.first] &= ~(CEPH_OSD_AUTOOUT | CEPH_OSD_NEW); | |
1653 | osd_xinfo[weight.first].old_weight = 0; | |
1654 | } | |
1655 | } | |
1656 | ||
1657 | for (const auto &primary_affinity : inc.new_primary_affinity) { | |
1658 | set_primary_affinity(primary_affinity.first, primary_affinity.second); | |
1659 | } | |
1660 | ||
1661 | // erasure_code_profiles | |
1662 | for (const auto &profile : inc.old_erasure_code_profiles) | |
1663 | erasure_code_profiles.erase(profile); | |
1664 | ||
1665 | for (const auto &profile : inc.new_erasure_code_profiles) { | |
1666 | set_erasure_code_profile(profile.first, profile.second); | |
1667 | } | |
1668 | ||
1669 | // up/down | |
1670 | for (const auto &state : inc.new_state) { | |
1671 | const auto osd = state.first; | |
1672 | int s = state.second ? state.second : CEPH_OSD_UP; | |
1673 | if ((osd_state[osd] & CEPH_OSD_UP) && | |
1674 | (s & CEPH_OSD_UP)) { | |
1675 | osd_info[osd].down_at = epoch; | |
1676 | osd_xinfo[osd].down_stamp = modified; | |
1677 | } | |
1678 | if ((osd_state[osd] & CEPH_OSD_EXISTS) && | |
1679 | (s & CEPH_OSD_EXISTS)) { | |
1680 | // osd is destroyed; clear out anything interesting. | |
1681 | (*osd_uuid)[osd] = uuid_d(); | |
1682 | osd_info[osd] = osd_info_t(); | |
1683 | osd_xinfo[osd] = osd_xinfo_t(); | |
1684 | set_primary_affinity(osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
1685 | osd_addrs->client_addr[osd].reset(new entity_addr_t()); | |
1686 | osd_addrs->cluster_addr[osd].reset(new entity_addr_t()); | |
1687 | osd_addrs->hb_front_addr[osd].reset(new entity_addr_t()); | |
1688 | osd_addrs->hb_back_addr[osd].reset(new entity_addr_t()); | |
1689 | osd_state[osd] = 0; | |
1690 | } else { | |
1691 | osd_state[osd] ^= s; | |
1692 | } | |
1693 | } | |
1694 | ||
1695 | for (const auto &client : inc.new_up_client) { | |
1696 | osd_state[client.first] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; | |
1697 | osd_addrs->client_addr[client.first].reset(new entity_addr_t(client.second)); | |
1698 | if (inc.new_hb_back_up.empty()) | |
1699 | osd_addrs->hb_back_addr[client.first].reset(new entity_addr_t(client.second)); //this is a backward-compatibility hack | |
1700 | else | |
1701 | osd_addrs->hb_back_addr[client.first].reset( | |
1702 | new entity_addr_t(inc.new_hb_back_up.find(client.first)->second)); | |
1703 | const auto j = inc.new_hb_front_up.find(client.first); | |
1704 | if (j != inc.new_hb_front_up.end()) | |
1705 | osd_addrs->hb_front_addr[client.first].reset(new entity_addr_t(j->second)); | |
1706 | else | |
1707 | osd_addrs->hb_front_addr[client.first].reset(); | |
1708 | ||
1709 | osd_info[client.first].up_from = epoch; | |
1710 | } | |
1711 | ||
1712 | for (const auto &cluster : inc.new_up_cluster) | |
1713 | osd_addrs->cluster_addr[cluster.first].reset(new entity_addr_t(cluster.second)); | |
1714 | ||
1715 | // info | |
1716 | for (const auto &thru : inc.new_up_thru) | |
1717 | osd_info[thru.first].up_thru = thru.second; | |
1718 | ||
1719 | for (const auto &interval : inc.new_last_clean_interval) { | |
1720 | osd_info[interval.first].last_clean_begin = interval.second.first; | |
1721 | osd_info[interval.first].last_clean_end = interval.second.second; | |
1722 | } | |
1723 | ||
1724 | for (const auto &lost : inc.new_lost) | |
1725 | osd_info[lost.first].lost_at = lost.second; | |
1726 | ||
1727 | // xinfo | |
1728 | for (const auto &xinfo : inc.new_xinfo) | |
1729 | osd_xinfo[xinfo.first] = xinfo.second; | |
1730 | ||
1731 | // uuid | |
1732 | for (const auto &uuid : inc.new_uuid) | |
1733 | (*osd_uuid)[uuid.first] = uuid.second; | |
1734 | ||
1735 | // pg rebuild | |
1736 | for (const auto &pg : inc.new_pg_temp) { | |
1737 | if (pg.second.empty()) | |
1738 | pg_temp->erase(pg.first); | |
1739 | else | |
31f18b77 FG |
1740 | pg_temp->set(pg.first, pg.second); |
1741 | } | |
1742 | if (!inc.new_pg_temp.empty()) { | |
1743 | // make sure pg_temp is efficiently stored | |
1744 | pg_temp->rebuild(); | |
7c673cae FG |
1745 | } |
1746 | ||
1747 | for (const auto &pg : inc.new_primary_temp) { | |
1748 | if (pg.second == -1) | |
1749 | primary_temp->erase(pg.first); | |
1750 | else | |
1751 | (*primary_temp)[pg.first] = pg.second; | |
1752 | } | |
1753 | ||
1754 | for (auto& p : inc.new_pg_upmap) { | |
1755 | pg_upmap[p.first] = p.second; | |
1756 | } | |
1757 | for (auto& pg : inc.old_pg_upmap) { | |
1758 | pg_upmap.erase(pg); | |
1759 | } | |
1760 | for (auto& p : inc.new_pg_upmap_items) { | |
1761 | pg_upmap_items[p.first] = p.second; | |
1762 | } | |
1763 | for (auto& pg : inc.old_pg_upmap_items) { | |
1764 | pg_upmap_items.erase(pg); | |
1765 | } | |
1766 | ||
1767 | // blacklist | |
1768 | if (!inc.new_blacklist.empty()) { | |
1769 | blacklist.insert(inc.new_blacklist.begin(),inc.new_blacklist.end()); | |
1770 | new_blacklist_entries = true; | |
1771 | } | |
1772 | for (const auto &addr : inc.old_blacklist) | |
1773 | blacklist.erase(addr); | |
1774 | ||
1775 | // cluster snapshot? | |
1776 | if (inc.cluster_snapshot.length()) { | |
1777 | cluster_snapshot = inc.cluster_snapshot; | |
1778 | cluster_snapshot_epoch = inc.epoch; | |
1779 | } else { | |
1780 | cluster_snapshot.clear(); | |
1781 | cluster_snapshot_epoch = 0; | |
1782 | } | |
1783 | ||
1784 | if (inc.new_nearfull_ratio >= 0) { | |
1785 | nearfull_ratio = inc.new_nearfull_ratio; | |
1786 | } | |
1787 | if (inc.new_backfillfull_ratio >= 0) { | |
1788 | backfillfull_ratio = inc.new_backfillfull_ratio; | |
1789 | } | |
1790 | if (inc.new_full_ratio >= 0) { | |
1791 | full_ratio = inc.new_full_ratio; | |
1792 | } | |
31f18b77 | 1793 | if (inc.new_require_min_compat_client > 0) { |
7c673cae FG |
1794 | require_min_compat_client = inc.new_require_min_compat_client; |
1795 | } | |
31f18b77 FG |
1796 | if (inc.new_require_osd_release >= 0) { |
1797 | require_osd_release = inc.new_require_osd_release; | |
1798 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { | |
1799 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 1800 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
1801 | } |
1802 | } | |
7c673cae FG |
1803 | |
1804 | // do new crush map last (after up/down stuff) | |
1805 | if (inc.crush.length()) { | |
1806 | bufferlist bl(inc.crush); | |
1807 | auto blp = bl.begin(); | |
1808 | crush.reset(new CrushWrapper); | |
1809 | crush->decode(blp); | |
31f18b77 FG |
1810 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { |
1811 | // only increment if this is a luminous-encoded osdmap, lest | |
1812 | // the mon's crush_version diverge from what the osds or others | |
1813 | // are decoding and applying on their end. if we won't encode | |
1814 | // it in the canonical version, don't change it. | |
1815 | ++crush_version; | |
1816 | } | |
7c673cae FG |
1817 | } |
1818 | ||
1819 | calc_num_osds(); | |
1820 | _calc_up_osd_features(); | |
1821 | return 0; | |
1822 | } | |
1823 | ||
1824 | // mapping | |
1825 | int OSDMap::map_to_pg( | |
1826 | int64_t poolid, | |
1827 | const string& name, | |
1828 | const string& key, | |
1829 | const string& nspace, | |
1830 | pg_t *pg) const | |
1831 | { | |
1832 | // calculate ps (placement seed) | |
1833 | const pg_pool_t *pool = get_pg_pool(poolid); | |
1834 | if (!pool) | |
1835 | return -ENOENT; | |
1836 | ps_t ps; | |
1837 | if (!key.empty()) | |
1838 | ps = pool->hash_key(key, nspace); | |
1839 | else | |
1840 | ps = pool->hash_key(name, nspace); | |
1841 | *pg = pg_t(ps, poolid); | |
1842 | return 0; | |
1843 | } | |
1844 | ||
1845 | int OSDMap::object_locator_to_pg( | |
1846 | const object_t& oid, const object_locator_t& loc, pg_t &pg) const | |
1847 | { | |
1848 | if (loc.hash >= 0) { | |
1849 | if (!get_pg_pool(loc.get_pool())) { | |
1850 | return -ENOENT; | |
1851 | } | |
1852 | pg = pg_t(loc.hash, loc.get_pool()); | |
1853 | return 0; | |
1854 | } | |
1855 | return map_to_pg(loc.get_pool(), oid.name, loc.key, loc.nspace, &pg); | |
1856 | } | |
1857 | ||
1858 | ceph_object_layout OSDMap::make_object_layout( | |
1859 | object_t oid, int pg_pool, string nspace) const | |
1860 | { | |
1861 | object_locator_t loc(pg_pool, nspace); | |
1862 | ||
1863 | ceph_object_layout ol; | |
1864 | pg_t pgid = object_locator_to_pg(oid, loc); | |
1865 | ol.ol_pgid = pgid.get_old_pg().v; | |
1866 | ol.ol_stripe_unit = 0; | |
1867 | return ol; | |
1868 | } | |
1869 | ||
1870 | void OSDMap::_remove_nonexistent_osds(const pg_pool_t& pool, | |
1871 | vector<int>& osds) const | |
1872 | { | |
1873 | if (pool.can_shift_osds()) { | |
1874 | unsigned removed = 0; | |
1875 | for (unsigned i = 0; i < osds.size(); i++) { | |
1876 | if (!exists(osds[i])) { | |
1877 | removed++; | |
1878 | continue; | |
1879 | } | |
1880 | if (removed) { | |
1881 | osds[i - removed] = osds[i]; | |
1882 | } | |
1883 | } | |
1884 | if (removed) | |
1885 | osds.resize(osds.size() - removed); | |
1886 | } else { | |
1887 | for (auto& osd : osds) { | |
1888 | if (!exists(osd)) | |
1889 | osd = CRUSH_ITEM_NONE; | |
1890 | } | |
1891 | } | |
1892 | } | |
1893 | ||
31f18b77 | 1894 | void OSDMap::_pg_to_raw_osds( |
7c673cae FG |
1895 | const pg_pool_t& pool, pg_t pg, |
1896 | vector<int> *osds, | |
1897 | ps_t *ppps) const | |
1898 | { | |
1899 | // map to osds[] | |
1900 | ps_t pps = pool.raw_pg_to_pps(pg); // placement ps | |
1901 | unsigned size = pool.get_size(); | |
1902 | ||
1903 | // what crush rule? | |
31f18b77 | 1904 | int ruleno = crush->find_rule(pool.get_crush_rule(), pool.get_type(), size); |
7c673cae FG |
1905 | if (ruleno >= 0) |
1906 | crush->do_rule(ruleno, pps, *osds, size, osd_weight, pg.pool()); | |
1907 | ||
1908 | _remove_nonexistent_osds(pool, *osds); | |
1909 | ||
1910 | if (ppps) | |
1911 | *ppps = pps; | |
7c673cae FG |
1912 | } |
1913 | ||
1914 | int OSDMap::_pick_primary(const vector<int>& osds) const | |
1915 | { | |
1916 | for (auto osd : osds) { | |
1917 | if (osd != CRUSH_ITEM_NONE) { | |
1918 | return osd; | |
1919 | } | |
1920 | } | |
1921 | return -1; | |
1922 | } | |
1923 | ||
224ce89b | 1924 | void OSDMap::_apply_upmap(const pg_pool_t& pi, pg_t raw_pg, vector<int> *raw) const |
7c673cae FG |
1925 | { |
1926 | pg_t pg = pi.raw_pg_to_pg(raw_pg); | |
1927 | auto p = pg_upmap.find(pg); | |
1928 | if (p != pg_upmap.end()) { | |
1929 | // make sure targets aren't marked out | |
1930 | for (auto osd : p->second) { | |
1931 | if (osd != CRUSH_ITEM_NONE && osd < max_osd && osd_weight[osd] == 0) { | |
1932 | // reject/ignore the explicit mapping | |
1933 | return; | |
1934 | } | |
1935 | } | |
1936 | *raw = vector<int>(p->second.begin(), p->second.end()); | |
224ce89b | 1937 | // continue to check and apply pg_upmap_items if any |
7c673cae FG |
1938 | } |
1939 | ||
1940 | auto q = pg_upmap_items.find(pg); | |
1941 | if (q != pg_upmap_items.end()) { | |
181888fb FG |
1942 | // NOTE: this approach does not allow a bidirectional swap, |
1943 | // e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1]. | |
1944 | for (auto& r : q->second) { | |
1945 | // make sure the replacement value doesn't already appear | |
1946 | bool exists = false; | |
1947 | ssize_t pos = -1; | |
1948 | for (unsigned i = 0; i < raw->size(); ++i) { | |
1949 | int osd = (*raw)[i]; | |
1950 | if (osd == r.second) { | |
1951 | exists = true; | |
1952 | break; | |
1953 | } | |
1954 | // ignore mapping if target is marked out (or invalid osd id) | |
1955 | if (osd == r.first && | |
1956 | pos < 0 && | |
1957 | !(r.second != CRUSH_ITEM_NONE && r.second < max_osd && | |
1958 | osd_weight[r.second] == 0)) { | |
1959 | pos = i; | |
1960 | } | |
1961 | } | |
1962 | if (!exists && pos >= 0) { | |
1963 | (*raw)[pos] = r.second; | |
7c673cae FG |
1964 | } |
1965 | } | |
1966 | } | |
1967 | } | |
1968 | ||
1969 | // pg -> (up osd list) | |
1970 | void OSDMap::_raw_to_up_osds(const pg_pool_t& pool, const vector<int>& raw, | |
1971 | vector<int> *up) const | |
1972 | { | |
1973 | if (pool.can_shift_osds()) { | |
1974 | // shift left | |
1975 | up->clear(); | |
1976 | up->reserve(raw.size()); | |
1977 | for (unsigned i=0; i<raw.size(); i++) { | |
1978 | if (!exists(raw[i]) || is_down(raw[i])) | |
1979 | continue; | |
1980 | up->push_back(raw[i]); | |
1981 | } | |
1982 | } else { | |
1983 | // set down/dne devices to NONE | |
1984 | up->resize(raw.size()); | |
1985 | for (int i = raw.size() - 1; i >= 0; --i) { | |
1986 | if (!exists(raw[i]) || is_down(raw[i])) { | |
1987 | (*up)[i] = CRUSH_ITEM_NONE; | |
1988 | } else { | |
1989 | (*up)[i] = raw[i]; | |
1990 | } | |
1991 | } | |
1992 | } | |
1993 | } | |
1994 | ||
1995 | void OSDMap::_apply_primary_affinity(ps_t seed, | |
1996 | const pg_pool_t& pool, | |
1997 | vector<int> *osds, | |
1998 | int *primary) const | |
1999 | { | |
2000 | // do we have any non-default primary_affinity values for these osds? | |
2001 | if (!osd_primary_affinity) | |
2002 | return; | |
2003 | ||
2004 | bool any = false; | |
2005 | for (const auto osd : *osds) { | |
2006 | if (osd != CRUSH_ITEM_NONE && | |
2007 | (*osd_primary_affinity)[osd] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
2008 | any = true; | |
2009 | break; | |
2010 | } | |
2011 | } | |
2012 | if (!any) | |
2013 | return; | |
2014 | ||
2015 | // pick the primary. feed both the seed (for the pg) and the osd | |
2016 | // into the hash/rng so that a proportional fraction of an osd's pgs | |
2017 | // get rejected as primary. | |
2018 | int pos = -1; | |
2019 | for (unsigned i = 0; i < osds->size(); ++i) { | |
2020 | int o = (*osds)[i]; | |
2021 | if (o == CRUSH_ITEM_NONE) | |
2022 | continue; | |
2023 | unsigned a = (*osd_primary_affinity)[o]; | |
2024 | if (a < CEPH_OSD_MAX_PRIMARY_AFFINITY && | |
2025 | (crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
2026 | seed, o) >> 16) >= a) { | |
2027 | // we chose not to use this primary. note it anyway as a | |
2028 | // fallback in case we don't pick anyone else, but keep looking. | |
2029 | if (pos < 0) | |
2030 | pos = i; | |
2031 | } else { | |
2032 | pos = i; | |
2033 | break; | |
2034 | } | |
2035 | } | |
2036 | if (pos < 0) | |
2037 | return; | |
2038 | ||
2039 | *primary = (*osds)[pos]; | |
2040 | ||
2041 | if (pool.can_shift_osds() && pos > 0) { | |
2042 | // move the new primary to the front. | |
2043 | for (int i = pos; i > 0; --i) { | |
2044 | (*osds)[i] = (*osds)[i-1]; | |
2045 | } | |
2046 | (*osds)[0] = *primary; | |
2047 | } | |
2048 | } | |
2049 | ||
2050 | void OSDMap::_get_temp_osds(const pg_pool_t& pool, pg_t pg, | |
2051 | vector<int> *temp_pg, int *temp_primary) const | |
2052 | { | |
2053 | pg = pool.raw_pg_to_pg(pg); | |
2054 | const auto p = pg_temp->find(pg); | |
2055 | temp_pg->clear(); | |
2056 | if (p != pg_temp->end()) { | |
2057 | for (unsigned i=0; i<p->second.size(); i++) { | |
2058 | if (!exists(p->second[i]) || is_down(p->second[i])) { | |
2059 | if (pool.can_shift_osds()) { | |
2060 | continue; | |
2061 | } else { | |
2062 | temp_pg->push_back(CRUSH_ITEM_NONE); | |
2063 | } | |
2064 | } else { | |
2065 | temp_pg->push_back(p->second[i]); | |
2066 | } | |
2067 | } | |
2068 | } | |
2069 | const auto &pp = primary_temp->find(pg); | |
2070 | *temp_primary = -1; | |
2071 | if (pp != primary_temp->end()) { | |
2072 | *temp_primary = pp->second; | |
2073 | } else if (!temp_pg->empty()) { // apply pg_temp's primary | |
2074 | for (unsigned i = 0; i < temp_pg->size(); ++i) { | |
2075 | if ((*temp_pg)[i] != CRUSH_ITEM_NONE) { | |
2076 | *temp_primary = (*temp_pg)[i]; | |
2077 | break; | |
2078 | } | |
2079 | } | |
2080 | } | |
2081 | } | |
2082 | ||
31f18b77 | 2083 | void OSDMap::pg_to_raw_osds(pg_t pg, vector<int> *raw, int *primary) const |
7c673cae FG |
2084 | { |
2085 | *primary = -1; | |
2086 | raw->clear(); | |
2087 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2088 | if (!pool) | |
31f18b77 FG |
2089 | return; |
2090 | _pg_to_raw_osds(*pool, pg, raw, NULL); | |
7c673cae FG |
2091 | if (primary) |
2092 | *primary = _pick_primary(*raw); | |
7c673cae FG |
2093 | } |
2094 | ||
2095 | void OSDMap::pg_to_raw_up(pg_t pg, vector<int> *up, int *primary) const | |
2096 | { | |
2097 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2098 | if (!pool) { | |
2099 | if (primary) | |
2100 | *primary = -1; | |
2101 | if (up) | |
2102 | up->clear(); | |
2103 | return; | |
2104 | } | |
2105 | vector<int> raw; | |
2106 | ps_t pps; | |
2107 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2108 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2109 | _raw_to_up_osds(*pool, raw, up); |
2110 | *primary = _pick_primary(raw); | |
2111 | _apply_primary_affinity(pps, *pool, up, primary); | |
2112 | } | |
31f18b77 | 2113 | |
7c673cae FG |
2114 | void OSDMap::_pg_to_up_acting_osds( |
2115 | const pg_t& pg, vector<int> *up, int *up_primary, | |
2116 | vector<int> *acting, int *acting_primary, | |
2117 | bool raw_pg_to_pg) const | |
2118 | { | |
2119 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2120 | if (!pool || | |
2121 | (!raw_pg_to_pg && pg.ps() >= pool->get_pg_num())) { | |
2122 | if (up) | |
2123 | up->clear(); | |
2124 | if (up_primary) | |
2125 | *up_primary = -1; | |
2126 | if (acting) | |
2127 | acting->clear(); | |
2128 | if (acting_primary) | |
2129 | *acting_primary = -1; | |
2130 | return; | |
2131 | } | |
2132 | vector<int> raw; | |
2133 | vector<int> _up; | |
2134 | vector<int> _acting; | |
2135 | int _up_primary; | |
2136 | int _acting_primary; | |
2137 | ps_t pps; | |
2138 | _get_temp_osds(*pool, pg, &_acting, &_acting_primary); | |
2139 | if (_acting.empty() || up || up_primary) { | |
2140 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2141 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2142 | _raw_to_up_osds(*pool, raw, &_up); |
2143 | _up_primary = _pick_primary(_up); | |
2144 | _apply_primary_affinity(pps, *pool, &_up, &_up_primary); | |
2145 | if (_acting.empty()) { | |
2146 | _acting = _up; | |
2147 | if (_acting_primary == -1) { | |
2148 | _acting_primary = _up_primary; | |
2149 | } | |
2150 | } | |
2151 | ||
2152 | if (up) | |
2153 | up->swap(_up); | |
2154 | if (up_primary) | |
2155 | *up_primary = _up_primary; | |
2156 | } | |
2157 | ||
2158 | if (acting) | |
2159 | acting->swap(_acting); | |
2160 | if (acting_primary) | |
2161 | *acting_primary = _acting_primary; | |
2162 | } | |
2163 | ||
2164 | int OSDMap::calc_pg_rank(int osd, const vector<int>& acting, int nrep) | |
2165 | { | |
2166 | if (!nrep) | |
2167 | nrep = acting.size(); | |
2168 | for (int i=0; i<nrep; i++) | |
2169 | if (acting[i] == osd) | |
2170 | return i; | |
2171 | return -1; | |
2172 | } | |
2173 | ||
2174 | int OSDMap::calc_pg_role(int osd, const vector<int>& acting, int nrep) | |
2175 | { | |
2176 | return calc_pg_rank(osd, acting, nrep); | |
2177 | } | |
2178 | ||
2179 | bool OSDMap::primary_changed( | |
2180 | int oldprimary, | |
2181 | const vector<int> &oldacting, | |
2182 | int newprimary, | |
2183 | const vector<int> &newacting) | |
2184 | { | |
2185 | if (oldacting.empty() && newacting.empty()) | |
2186 | return false; // both still empty | |
2187 | if (oldacting.empty() ^ newacting.empty()) | |
2188 | return true; // was empty, now not, or vice versa | |
2189 | if (oldprimary != newprimary) | |
2190 | return true; // primary changed | |
2191 | if (calc_pg_rank(oldprimary, oldacting) != | |
2192 | calc_pg_rank(newprimary, newacting)) | |
2193 | return true; | |
2194 | return false; // same primary (tho replicas may have changed) | |
2195 | } | |
2196 | ||
2197 | ||
2198 | // serialize, unserialize | |
2199 | void OSDMap::encode_client_old(bufferlist& bl) const | |
2200 | { | |
2201 | __u16 v = 5; | |
2202 | ::encode(v, bl); | |
2203 | ||
2204 | // base | |
2205 | ::encode(fsid, bl); | |
2206 | ::encode(epoch, bl); | |
2207 | ::encode(created, bl); | |
2208 | ::encode(modified, bl); | |
2209 | ||
2210 | // for ::encode(pools, bl); | |
2211 | __u32 n = pools.size(); | |
2212 | ::encode(n, bl); | |
2213 | ||
2214 | for (const auto &pool : pools) { | |
2215 | n = pool.first; | |
2216 | ::encode(n, bl); | |
2217 | ::encode(pool.second, bl, 0); | |
2218 | } | |
2219 | // for ::encode(pool_name, bl); | |
2220 | n = pool_name.size(); | |
2221 | ::encode(n, bl); | |
2222 | for (const auto &pname : pool_name) { | |
2223 | n = pname.first; | |
2224 | ::encode(n, bl); | |
2225 | ::encode(pname.second, bl); | |
2226 | } | |
2227 | // for ::encode(pool_max, bl); | |
2228 | n = pool_max; | |
2229 | ::encode(n, bl); | |
2230 | ||
2231 | ::encode(flags, bl); | |
2232 | ||
2233 | ::encode(max_osd, bl); | |
31f18b77 FG |
2234 | { |
2235 | uint32_t n = osd_state.size(); | |
2236 | ::encode(n, bl); | |
2237 | for (auto s : osd_state) { | |
2238 | ::encode((uint8_t)s, bl); | |
2239 | } | |
2240 | } | |
7c673cae FG |
2241 | ::encode(osd_weight, bl); |
2242 | ::encode(osd_addrs->client_addr, bl, 0); | |
2243 | ||
2244 | // for ::encode(pg_temp, bl); | |
2245 | n = pg_temp->size(); | |
2246 | ::encode(n, bl); | |
2247 | for (const auto pg : *pg_temp) { | |
2248 | old_pg_t opg = pg.first.get_old_pg(); | |
2249 | ::encode(opg, bl); | |
2250 | ::encode(pg.second, bl); | |
2251 | } | |
2252 | ||
2253 | // crush | |
2254 | bufferlist cbl; | |
2255 | crush->encode(cbl, 0 /* legacy (no) features */); | |
2256 | ::encode(cbl, bl); | |
2257 | } | |
2258 | ||
2259 | void OSDMap::encode_classic(bufferlist& bl, uint64_t features) const | |
2260 | { | |
2261 | if ((features & CEPH_FEATURE_PGID64) == 0) { | |
2262 | encode_client_old(bl); | |
2263 | return; | |
2264 | } | |
2265 | ||
2266 | __u16 v = 6; | |
2267 | ::encode(v, bl); | |
2268 | ||
2269 | // base | |
2270 | ::encode(fsid, bl); | |
2271 | ::encode(epoch, bl); | |
2272 | ::encode(created, bl); | |
2273 | ::encode(modified, bl); | |
2274 | ||
2275 | ::encode(pools, bl, features); | |
2276 | ::encode(pool_name, bl); | |
2277 | ::encode(pool_max, bl); | |
2278 | ||
2279 | ::encode(flags, bl); | |
2280 | ||
2281 | ::encode(max_osd, bl); | |
31f18b77 FG |
2282 | { |
2283 | uint32_t n = osd_state.size(); | |
2284 | ::encode(n, bl); | |
2285 | for (auto s : osd_state) { | |
2286 | ::encode((uint8_t)s, bl); | |
2287 | } | |
2288 | } | |
7c673cae FG |
2289 | ::encode(osd_weight, bl); |
2290 | ::encode(osd_addrs->client_addr, bl, features); | |
2291 | ||
2292 | ::encode(*pg_temp, bl); | |
2293 | ||
2294 | // crush | |
2295 | bufferlist cbl; | |
2296 | crush->encode(cbl, 0 /* legacy (no) features */); | |
2297 | ::encode(cbl, bl); | |
2298 | ||
2299 | // extended | |
2300 | __u16 ev = 10; | |
2301 | ::encode(ev, bl); | |
2302 | ::encode(osd_addrs->hb_back_addr, bl, features); | |
2303 | ::encode(osd_info, bl); | |
2304 | ::encode(blacklist, bl, features); | |
2305 | ::encode(osd_addrs->cluster_addr, bl, features); | |
2306 | ::encode(cluster_snapshot_epoch, bl); | |
2307 | ::encode(cluster_snapshot, bl); | |
2308 | ::encode(*osd_uuid, bl); | |
2309 | ::encode(osd_xinfo, bl); | |
2310 | ::encode(osd_addrs->hb_front_addr, bl, features); | |
2311 | } | |
2312 | ||
2313 | void OSDMap::encode(bufferlist& bl, uint64_t features) const | |
2314 | { | |
2315 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { | |
2316 | encode_classic(bl, features); | |
2317 | return; | |
2318 | } | |
2319 | ||
2320 | // only a select set of callers should *ever* be encoding new | |
2321 | // OSDMaps. others should be passing around the canonical encoded | |
2322 | // buffers from on high. select out those callers by passing in an | |
2323 | // "impossible" feature bit. | |
2324 | assert(features & CEPH_FEATURE_RESERVED); | |
2325 | features &= ~CEPH_FEATURE_RESERVED; | |
2326 | ||
2327 | size_t start_offset = bl.length(); | |
2328 | size_t tail_offset; | |
2329 | buffer::list::iterator crc_it; | |
2330 | ||
2331 | // meta-encoding: how we include client-used and osd-specific data | |
2332 | ENCODE_START(8, 7, bl); | |
2333 | ||
2334 | { | |
31f18b77 FG |
2335 | uint8_t v = 6; |
2336 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { | |
7c673cae FG |
2337 | v = 3; |
2338 | } | |
2339 | ENCODE_START(v, 1, bl); // client-usable data | |
2340 | // base | |
2341 | ::encode(fsid, bl); | |
2342 | ::encode(epoch, bl); | |
2343 | ::encode(created, bl); | |
2344 | ::encode(modified, bl); | |
2345 | ||
2346 | ::encode(pools, bl, features); | |
2347 | ::encode(pool_name, bl); | |
2348 | ::encode(pool_max, bl); | |
2349 | ||
31f18b77 FG |
2350 | if (v < 4) { |
2351 | decltype(flags) f = flags; | |
2352 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) | |
c07f9fc5 | 2353 | f |= CEPH_OSDMAP_REQUIRE_LUMINOUS | CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2354 | else if (require_osd_release == CEPH_RELEASE_KRAKEN) |
2355 | f |= CEPH_OSDMAP_REQUIRE_KRAKEN; | |
2356 | else if (require_osd_release == CEPH_RELEASE_JEWEL) | |
2357 | f |= CEPH_OSDMAP_REQUIRE_JEWEL; | |
2358 | ::encode(f, bl); | |
2359 | } else { | |
2360 | ::encode(flags, bl); | |
2361 | } | |
7c673cae FG |
2362 | |
2363 | ::encode(max_osd, bl); | |
31f18b77 FG |
2364 | if (v >= 5) { |
2365 | ::encode(osd_state, bl); | |
2366 | } else { | |
2367 | uint32_t n = osd_state.size(); | |
2368 | ::encode(n, bl); | |
2369 | for (auto s : osd_state) { | |
2370 | ::encode((uint8_t)s, bl); | |
2371 | } | |
2372 | } | |
7c673cae FG |
2373 | ::encode(osd_weight, bl); |
2374 | ::encode(osd_addrs->client_addr, bl, features); | |
2375 | ||
2376 | ::encode(*pg_temp, bl); | |
2377 | ::encode(*primary_temp, bl); | |
2378 | if (osd_primary_affinity) { | |
2379 | ::encode(*osd_primary_affinity, bl); | |
2380 | } else { | |
2381 | vector<__u32> v; | |
2382 | ::encode(v, bl); | |
2383 | } | |
2384 | ||
2385 | // crush | |
2386 | bufferlist cbl; | |
2387 | crush->encode(cbl, features); | |
2388 | ::encode(cbl, bl); | |
2389 | ::encode(erasure_code_profiles, bl); | |
2390 | ||
2391 | if (v >= 4) { | |
2392 | ::encode(pg_upmap, bl); | |
2393 | ::encode(pg_upmap_items, bl); | |
2394 | } else { | |
2395 | assert(pg_upmap.empty()); | |
2396 | assert(pg_upmap_items.empty()); | |
2397 | } | |
31f18b77 FG |
2398 | if (v >= 6) { |
2399 | ::encode(crush_version, bl); | |
2400 | } | |
7c673cae FG |
2401 | ENCODE_FINISH(bl); // client-usable data |
2402 | } | |
2403 | ||
2404 | { | |
31f18b77 | 2405 | uint8_t target_v = 5; |
7c673cae FG |
2406 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
2407 | target_v = 1; | |
2408 | } | |
2409 | ENCODE_START(target_v, 1, bl); // extended, osd-only data | |
2410 | ::encode(osd_addrs->hb_back_addr, bl, features); | |
2411 | ::encode(osd_info, bl); | |
2412 | { | |
2413 | // put this in a sorted, ordered map<> so that we encode in a | |
2414 | // deterministic order. | |
2415 | map<entity_addr_t,utime_t> blacklist_map; | |
2416 | for (const auto &addr : blacklist) | |
2417 | blacklist_map.insert(make_pair(addr.first, addr.second)); | |
2418 | ::encode(blacklist_map, bl, features); | |
2419 | } | |
2420 | ::encode(osd_addrs->cluster_addr, bl, features); | |
2421 | ::encode(cluster_snapshot_epoch, bl); | |
2422 | ::encode(cluster_snapshot, bl); | |
2423 | ::encode(*osd_uuid, bl); | |
2424 | ::encode(osd_xinfo, bl); | |
2425 | ::encode(osd_addrs->hb_front_addr, bl, features); | |
2426 | if (target_v >= 2) { | |
2427 | ::encode(nearfull_ratio, bl); | |
2428 | ::encode(full_ratio, bl); | |
2429 | ::encode(backfillfull_ratio, bl); | |
31f18b77 FG |
2430 | } |
2431 | // 4 was string-based new_require_min_compat_client | |
2432 | if (target_v >= 5) { | |
7c673cae | 2433 | ::encode(require_min_compat_client, bl); |
31f18b77 | 2434 | ::encode(require_osd_release, bl); |
7c673cae FG |
2435 | } |
2436 | ENCODE_FINISH(bl); // osd-only data | |
2437 | } | |
2438 | ||
2439 | ::encode((uint32_t)0, bl); // dummy crc | |
2440 | crc_it = bl.end(); | |
2441 | crc_it.advance(-4); | |
2442 | tail_offset = bl.length(); | |
2443 | ||
2444 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
2445 | ||
2446 | // fill in crc | |
2447 | bufferlist front; | |
2448 | front.substr_of(bl, start_offset, crc_it.get_off() - start_offset); | |
2449 | crc = front.crc32c(-1); | |
2450 | if (tail_offset < bl.length()) { | |
2451 | bufferlist tail; | |
2452 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); | |
2453 | crc = tail.crc32c(crc); | |
2454 | } | |
2455 | ceph_le32 crc_le; | |
2456 | crc_le = crc; | |
2457 | crc_it.copy_in(4, (char*)&crc_le); | |
2458 | crc_defined = true; | |
2459 | } | |
2460 | ||
2461 | void OSDMap::decode(bufferlist& bl) | |
2462 | { | |
2463 | auto p = bl.begin(); | |
2464 | decode(p); | |
2465 | } | |
2466 | ||
2467 | void OSDMap::decode_classic(bufferlist::iterator& p) | |
2468 | { | |
2469 | __u32 n, t; | |
2470 | __u16 v; | |
2471 | ::decode(v, p); | |
2472 | ||
2473 | // base | |
2474 | ::decode(fsid, p); | |
2475 | ::decode(epoch, p); | |
2476 | ::decode(created, p); | |
2477 | ::decode(modified, p); | |
2478 | ||
2479 | if (v < 6) { | |
2480 | if (v < 4) { | |
2481 | int32_t max_pools = 0; | |
2482 | ::decode(max_pools, p); | |
2483 | pool_max = max_pools; | |
2484 | } | |
2485 | pools.clear(); | |
2486 | ::decode(n, p); | |
2487 | while (n--) { | |
2488 | ::decode(t, p); | |
2489 | ::decode(pools[t], p); | |
2490 | } | |
2491 | if (v == 4) { | |
2492 | ::decode(n, p); | |
2493 | pool_max = n; | |
2494 | } else if (v == 5) { | |
2495 | pool_name.clear(); | |
2496 | ::decode(n, p); | |
2497 | while (n--) { | |
2498 | ::decode(t, p); | |
2499 | ::decode(pool_name[t], p); | |
2500 | } | |
2501 | ::decode(n, p); | |
2502 | pool_max = n; | |
2503 | } | |
2504 | } else { | |
2505 | ::decode(pools, p); | |
2506 | ::decode(pool_name, p); | |
2507 | ::decode(pool_max, p); | |
2508 | } | |
2509 | // kludge around some old bug that zeroed out pool_max (#2307) | |
2510 | if (pools.size() && pool_max < pools.rbegin()->first) { | |
2511 | pool_max = pools.rbegin()->first; | |
2512 | } | |
2513 | ||
2514 | ::decode(flags, p); | |
2515 | ||
2516 | ::decode(max_osd, p); | |
31f18b77 FG |
2517 | { |
2518 | vector<uint8_t> os; | |
2519 | ::decode(os, p); | |
2520 | osd_state.resize(os.size()); | |
2521 | for (unsigned i = 0; i < os.size(); ++i) { | |
2522 | osd_state[i] = os[i]; | |
2523 | } | |
2524 | } | |
7c673cae FG |
2525 | ::decode(osd_weight, p); |
2526 | ::decode(osd_addrs->client_addr, p); | |
2527 | if (v <= 5) { | |
2528 | pg_temp->clear(); | |
2529 | ::decode(n, p); | |
2530 | while (n--) { | |
2531 | old_pg_t opg; | |
2532 | ::decode_raw(opg, p); | |
31f18b77 FG |
2533 | mempool::osdmap::vector<int32_t> v; |
2534 | ::decode(v, p); | |
2535 | pg_temp->set(pg_t(opg), v); | |
7c673cae FG |
2536 | } |
2537 | } else { | |
2538 | ::decode(*pg_temp, p); | |
2539 | } | |
2540 | ||
2541 | // crush | |
2542 | bufferlist cbl; | |
2543 | ::decode(cbl, p); | |
2544 | auto cblp = cbl.begin(); | |
2545 | crush->decode(cblp); | |
2546 | ||
2547 | // extended | |
2548 | __u16 ev = 0; | |
2549 | if (v >= 5) | |
2550 | ::decode(ev, p); | |
2551 | ::decode(osd_addrs->hb_back_addr, p); | |
2552 | ::decode(osd_info, p); | |
2553 | if (v < 5) | |
2554 | ::decode(pool_name, p); | |
2555 | ||
2556 | ::decode(blacklist, p); | |
2557 | if (ev >= 6) | |
2558 | ::decode(osd_addrs->cluster_addr, p); | |
2559 | else | |
2560 | osd_addrs->cluster_addr.resize(osd_addrs->client_addr.size()); | |
2561 | ||
2562 | if (ev >= 7) { | |
2563 | ::decode(cluster_snapshot_epoch, p); | |
2564 | ::decode(cluster_snapshot, p); | |
2565 | } | |
2566 | ||
2567 | if (ev >= 8) { | |
2568 | ::decode(*osd_uuid, p); | |
2569 | } else { | |
2570 | osd_uuid->resize(max_osd); | |
2571 | } | |
2572 | if (ev >= 9) | |
2573 | ::decode(osd_xinfo, p); | |
2574 | else | |
2575 | osd_xinfo.resize(max_osd); | |
2576 | ||
2577 | if (ev >= 10) | |
2578 | ::decode(osd_addrs->hb_front_addr, p); | |
2579 | else | |
2580 | osd_addrs->hb_front_addr.resize(osd_addrs->hb_back_addr.size()); | |
2581 | ||
2582 | osd_primary_affinity.reset(); | |
2583 | ||
2584 | post_decode(); | |
2585 | } | |
2586 | ||
2587 | void OSDMap::decode(bufferlist::iterator& bl) | |
2588 | { | |
2589 | /** | |
2590 | * Older encodings of the OSDMap had a single struct_v which | |
2591 | * covered the whole encoding, and was prior to our modern | |
2592 | * stuff which includes a compatv and a size. So if we see | |
2593 | * a struct_v < 7, we must rewind to the beginning and use our | |
2594 | * classic decoder. | |
2595 | */ | |
2596 | size_t start_offset = bl.get_off(); | |
2597 | size_t tail_offset = 0; | |
2598 | bufferlist crc_front, crc_tail; | |
2599 | ||
2600 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
2601 | if (struct_v < 7) { | |
2602 | int struct_v_size = sizeof(struct_v); | |
2603 | bl.advance(-struct_v_size); | |
2604 | decode_classic(bl); | |
2605 | return; | |
2606 | } | |
2607 | /** | |
2608 | * Since we made it past that hurdle, we can use our normal paths. | |
2609 | */ | |
2610 | { | |
31f18b77 | 2611 | DECODE_START(6, bl); // client-usable data |
7c673cae FG |
2612 | // base |
2613 | ::decode(fsid, bl); | |
2614 | ::decode(epoch, bl); | |
2615 | ::decode(created, bl); | |
2616 | ::decode(modified, bl); | |
2617 | ||
2618 | ::decode(pools, bl); | |
2619 | ::decode(pool_name, bl); | |
2620 | ::decode(pool_max, bl); | |
2621 | ||
2622 | ::decode(flags, bl); | |
2623 | ||
2624 | ::decode(max_osd, bl); | |
31f18b77 FG |
2625 | if (struct_v >= 5) { |
2626 | ::decode(osd_state, bl); | |
2627 | } else { | |
2628 | vector<uint8_t> os; | |
2629 | ::decode(os, bl); | |
2630 | osd_state.resize(os.size()); | |
2631 | for (unsigned i = 0; i < os.size(); ++i) { | |
2632 | osd_state[i] = os[i]; | |
2633 | } | |
2634 | } | |
7c673cae FG |
2635 | ::decode(osd_weight, bl); |
2636 | ::decode(osd_addrs->client_addr, bl); | |
2637 | ||
2638 | ::decode(*pg_temp, bl); | |
2639 | ::decode(*primary_temp, bl); | |
2640 | if (struct_v >= 2) { | |
2641 | osd_primary_affinity.reset(new mempool::osdmap::vector<__u32>); | |
2642 | ::decode(*osd_primary_affinity, bl); | |
2643 | if (osd_primary_affinity->empty()) | |
2644 | osd_primary_affinity.reset(); | |
2645 | } else { | |
2646 | osd_primary_affinity.reset(); | |
2647 | } | |
2648 | ||
2649 | // crush | |
2650 | bufferlist cbl; | |
2651 | ::decode(cbl, bl); | |
2652 | auto cblp = cbl.begin(); | |
2653 | crush->decode(cblp); | |
2654 | if (struct_v >= 3) { | |
2655 | ::decode(erasure_code_profiles, bl); | |
2656 | } else { | |
2657 | erasure_code_profiles.clear(); | |
2658 | } | |
2659 | if (struct_v >= 4) { | |
2660 | ::decode(pg_upmap, bl); | |
2661 | ::decode(pg_upmap_items, bl); | |
2662 | } else { | |
2663 | pg_upmap.clear(); | |
2664 | pg_upmap_items.clear(); | |
2665 | } | |
31f18b77 FG |
2666 | if (struct_v >= 6) { |
2667 | ::decode(crush_version, bl); | |
2668 | } | |
7c673cae FG |
2669 | DECODE_FINISH(bl); // client-usable data |
2670 | } | |
2671 | ||
2672 | { | |
31f18b77 | 2673 | DECODE_START(5, bl); // extended, osd-only data |
7c673cae FG |
2674 | ::decode(osd_addrs->hb_back_addr, bl); |
2675 | ::decode(osd_info, bl); | |
2676 | ::decode(blacklist, bl); | |
2677 | ::decode(osd_addrs->cluster_addr, bl); | |
2678 | ::decode(cluster_snapshot_epoch, bl); | |
2679 | ::decode(cluster_snapshot, bl); | |
2680 | ::decode(*osd_uuid, bl); | |
2681 | ::decode(osd_xinfo, bl); | |
2682 | ::decode(osd_addrs->hb_front_addr, bl); | |
2683 | if (struct_v >= 2) { | |
2684 | ::decode(nearfull_ratio, bl); | |
2685 | ::decode(full_ratio, bl); | |
2686 | } else { | |
2687 | nearfull_ratio = 0; | |
2688 | full_ratio = 0; | |
2689 | } | |
2690 | if (struct_v >= 3) { | |
2691 | ::decode(backfillfull_ratio, bl); | |
2692 | } else { | |
2693 | backfillfull_ratio = 0; | |
2694 | } | |
31f18b77 FG |
2695 | if (struct_v == 4) { |
2696 | string r; | |
2697 | ::decode(r, bl); | |
2698 | if (r.length()) | |
2699 | require_min_compat_client = ceph_release_from_name(r.c_str()); | |
2700 | } | |
2701 | if (struct_v >= 5) { | |
7c673cae | 2702 | ::decode(require_min_compat_client, bl); |
31f18b77 FG |
2703 | ::decode(require_osd_release, bl); |
2704 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { | |
2705 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 2706 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2707 | } |
2708 | } else { | |
2709 | if (flags & CEPH_OSDMAP_REQUIRE_LUMINOUS) { | |
2710 | // only for compat with post-kraken pre-luminous test clusters | |
2711 | require_osd_release = CEPH_RELEASE_LUMINOUS; | |
2712 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 2713 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2714 | } else if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { |
2715 | require_osd_release = CEPH_RELEASE_KRAKEN; | |
2716 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
2717 | require_osd_release = CEPH_RELEASE_JEWEL; | |
2718 | } else { | |
2719 | require_osd_release = 0; | |
2720 | } | |
2721 | } | |
7c673cae FG |
2722 | DECODE_FINISH(bl); // osd-only data |
2723 | } | |
2724 | ||
2725 | if (struct_v >= 8) { | |
2726 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
2727 | ::decode(crc, bl); | |
2728 | tail_offset = bl.get_off(); | |
2729 | crc_defined = true; | |
2730 | } else { | |
2731 | crc_defined = false; | |
2732 | crc = 0; | |
2733 | } | |
2734 | ||
2735 | DECODE_FINISH(bl); // wrapper | |
2736 | ||
2737 | if (tail_offset) { | |
2738 | // verify crc | |
2739 | uint32_t actual = crc_front.crc32c(-1); | |
2740 | if (tail_offset < bl.get_off()) { | |
2741 | bufferlist tail; | |
2742 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); | |
2743 | actual = tail.crc32c(actual); | |
2744 | } | |
2745 | if (crc != actual) { | |
2746 | ostringstream ss; | |
2747 | ss << "bad crc, actual " << actual << " != expected " << crc; | |
2748 | string s = ss.str(); | |
2749 | throw buffer::malformed_input(s.c_str()); | |
2750 | } | |
2751 | } | |
2752 | ||
2753 | post_decode(); | |
2754 | } | |
2755 | ||
2756 | void OSDMap::post_decode() | |
2757 | { | |
2758 | // index pool names | |
2759 | name_pool.clear(); | |
2760 | for (const auto &pname : pool_name) { | |
2761 | name_pool[pname.second] = pname.first; | |
2762 | } | |
2763 | ||
2764 | calc_num_osds(); | |
2765 | _calc_up_osd_features(); | |
2766 | } | |
2767 | ||
2768 | void OSDMap::dump_erasure_code_profiles( | |
2769 | const mempool::osdmap::map<string,map<string,string>>& profiles, | |
2770 | Formatter *f) | |
2771 | { | |
2772 | f->open_object_section("erasure_code_profiles"); | |
2773 | for (const auto &profile : profiles) { | |
2774 | f->open_object_section(profile.first.c_str()); | |
2775 | for (const auto &profm : profile.second) { | |
2776 | f->dump_string(profm.first.c_str(), profm.second.c_str()); | |
2777 | } | |
2778 | f->close_section(); | |
2779 | } | |
2780 | f->close_section(); | |
2781 | } | |
2782 | ||
2783 | void OSDMap::dump(Formatter *f) const | |
2784 | { | |
2785 | f->dump_int("epoch", get_epoch()); | |
2786 | f->dump_stream("fsid") << get_fsid(); | |
2787 | f->dump_stream("created") << get_created(); | |
2788 | f->dump_stream("modified") << get_modified(); | |
2789 | f->dump_string("flags", get_flag_string()); | |
31f18b77 | 2790 | f->dump_unsigned("crush_version", get_crush_version()); |
7c673cae FG |
2791 | f->dump_float("full_ratio", full_ratio); |
2792 | f->dump_float("backfillfull_ratio", backfillfull_ratio); | |
2793 | f->dump_float("nearfull_ratio", nearfull_ratio); | |
2794 | f->dump_string("cluster_snapshot", get_cluster_snapshot()); | |
2795 | f->dump_int("pool_max", get_pool_max()); | |
2796 | f->dump_int("max_osd", get_max_osd()); | |
31f18b77 FG |
2797 | f->dump_string("require_min_compat_client", |
2798 | ceph_release_name(require_min_compat_client)); | |
2799 | f->dump_string("min_compat_client", | |
2800 | ceph_release_name(get_min_compat_client())); | |
2801 | f->dump_string("require_osd_release", | |
2802 | ceph_release_name(require_osd_release)); | |
7c673cae FG |
2803 | |
2804 | f->open_array_section("pools"); | |
2805 | for (const auto &pool : pools) { | |
2806 | std::string name("<unknown>"); | |
2807 | const auto &pni = pool_name.find(pool.first); | |
2808 | if (pni != pool_name.end()) | |
2809 | name = pni->second; | |
2810 | f->open_object_section("pool"); | |
2811 | f->dump_int("pool", pool.first); | |
2812 | f->dump_string("pool_name", name); | |
2813 | pool.second.dump(f); | |
2814 | f->close_section(); | |
2815 | } | |
2816 | f->close_section(); | |
2817 | ||
2818 | f->open_array_section("osds"); | |
2819 | for (int i=0; i<get_max_osd(); i++) | |
2820 | if (exists(i)) { | |
2821 | f->open_object_section("osd_info"); | |
2822 | f->dump_int("osd", i); | |
2823 | f->dump_stream("uuid") << get_uuid(i); | |
2824 | f->dump_int("up", is_up(i)); | |
2825 | f->dump_int("in", is_in(i)); | |
2826 | f->dump_float("weight", get_weightf(i)); | |
2827 | f->dump_float("primary_affinity", get_primary_affinityf(i)); | |
2828 | get_info(i).dump(f); | |
2829 | f->dump_stream("public_addr") << get_addr(i); | |
2830 | f->dump_stream("cluster_addr") << get_cluster_addr(i); | |
2831 | f->dump_stream("heartbeat_back_addr") << get_hb_back_addr(i); | |
2832 | f->dump_stream("heartbeat_front_addr") << get_hb_front_addr(i); | |
2833 | ||
2834 | set<string> st; | |
2835 | get_state(i, st); | |
2836 | f->open_array_section("state"); | |
2837 | for (const auto &state : st) | |
2838 | f->dump_string("state", state); | |
2839 | f->close_section(); | |
2840 | ||
2841 | f->close_section(); | |
2842 | } | |
2843 | f->close_section(); | |
2844 | ||
2845 | f->open_array_section("osd_xinfo"); | |
2846 | for (int i=0; i<get_max_osd(); i++) { | |
2847 | if (exists(i)) { | |
2848 | f->open_object_section("xinfo"); | |
2849 | f->dump_int("osd", i); | |
2850 | osd_xinfo[i].dump(f); | |
2851 | f->close_section(); | |
2852 | } | |
2853 | } | |
2854 | f->close_section(); | |
2855 | ||
2856 | f->open_array_section("pg_upmap"); | |
2857 | for (auto& p : pg_upmap) { | |
2858 | f->open_object_section("mapping"); | |
2859 | f->dump_stream("pgid") << p.first; | |
2860 | f->open_array_section("osds"); | |
2861 | for (auto q : p.second) { | |
2862 | f->dump_int("osd", q); | |
2863 | } | |
2864 | f->close_section(); | |
2865 | f->close_section(); | |
2866 | } | |
2867 | f->close_section(); | |
2868 | f->open_array_section("pg_upmap_items"); | |
2869 | for (auto& p : pg_upmap_items) { | |
2870 | f->open_object_section("mapping"); | |
2871 | f->dump_stream("pgid") << p.first; | |
2872 | f->open_array_section("mappings"); | |
2873 | for (auto& q : p.second) { | |
2874 | f->open_object_section("mapping"); | |
2875 | f->dump_int("from", q.first); | |
2876 | f->dump_int("to", q.second); | |
2877 | f->close_section(); | |
2878 | } | |
2879 | f->close_section(); | |
2880 | f->close_section(); | |
2881 | } | |
2882 | f->close_section(); | |
2883 | f->open_array_section("pg_temp"); | |
31f18b77 | 2884 | pg_temp->dump(f); |
7c673cae FG |
2885 | f->close_section(); |
2886 | ||
2887 | f->open_array_section("primary_temp"); | |
2888 | for (const auto &pg : *primary_temp) { | |
2889 | f->dump_stream("pgid") << pg.first; | |
2890 | f->dump_int("osd", pg.second); | |
2891 | } | |
2892 | f->close_section(); // primary_temp | |
2893 | ||
2894 | f->open_object_section("blacklist"); | |
2895 | for (const auto &addr : blacklist) { | |
2896 | stringstream ss; | |
2897 | ss << addr.first; | |
2898 | f->dump_stream(ss.str().c_str()) << addr.second; | |
2899 | } | |
2900 | f->close_section(); | |
2901 | ||
2902 | dump_erasure_code_profiles(erasure_code_profiles, f); | |
2903 | } | |
2904 | ||
2905 | void OSDMap::generate_test_instances(list<OSDMap*>& o) | |
2906 | { | |
2907 | o.push_back(new OSDMap); | |
2908 | ||
2909 | CephContext *cct = new CephContext(CODE_ENVIRONMENT_UTILITY); | |
2910 | o.push_back(new OSDMap); | |
2911 | uuid_d fsid; | |
224ce89b | 2912 | o.back()->build_simple(cct, 1, fsid, 16); |
7c673cae FG |
2913 | o.back()->created = o.back()->modified = utime_t(1, 2); // fix timestamp |
2914 | o.back()->blacklist[entity_addr_t()] = utime_t(5, 6); | |
2915 | cct->put(); | |
2916 | } | |
2917 | ||
2918 | string OSDMap::get_flag_string(unsigned f) | |
2919 | { | |
2920 | string s; | |
2921 | if ( f& CEPH_OSDMAP_NEARFULL) | |
2922 | s += ",nearfull"; | |
2923 | if (f & CEPH_OSDMAP_FULL) | |
2924 | s += ",full"; | |
2925 | if (f & CEPH_OSDMAP_PAUSERD) | |
2926 | s += ",pauserd"; | |
2927 | if (f & CEPH_OSDMAP_PAUSEWR) | |
2928 | s += ",pausewr"; | |
2929 | if (f & CEPH_OSDMAP_PAUSEREC) | |
2930 | s += ",pauserec"; | |
2931 | if (f & CEPH_OSDMAP_NOUP) | |
2932 | s += ",noup"; | |
2933 | if (f & CEPH_OSDMAP_NODOWN) | |
2934 | s += ",nodown"; | |
2935 | if (f & CEPH_OSDMAP_NOOUT) | |
2936 | s += ",noout"; | |
2937 | if (f & CEPH_OSDMAP_NOIN) | |
2938 | s += ",noin"; | |
2939 | if (f & CEPH_OSDMAP_NOBACKFILL) | |
2940 | s += ",nobackfill"; | |
2941 | if (f & CEPH_OSDMAP_NOREBALANCE) | |
2942 | s += ",norebalance"; | |
2943 | if (f & CEPH_OSDMAP_NORECOVER) | |
2944 | s += ",norecover"; | |
2945 | if (f & CEPH_OSDMAP_NOSCRUB) | |
2946 | s += ",noscrub"; | |
2947 | if (f & CEPH_OSDMAP_NODEEP_SCRUB) | |
2948 | s += ",nodeep-scrub"; | |
2949 | if (f & CEPH_OSDMAP_NOTIERAGENT) | |
2950 | s += ",notieragent"; | |
2951 | if (f & CEPH_OSDMAP_SORTBITWISE) | |
2952 | s += ",sortbitwise"; | |
2953 | if (f & CEPH_OSDMAP_REQUIRE_JEWEL) | |
2954 | s += ",require_jewel_osds"; | |
2955 | if (f & CEPH_OSDMAP_REQUIRE_KRAKEN) | |
2956 | s += ",require_kraken_osds"; | |
2957 | if (f & CEPH_OSDMAP_REQUIRE_LUMINOUS) | |
2958 | s += ",require_luminous_osds"; | |
c07f9fc5 FG |
2959 | if (f & CEPH_OSDMAP_RECOVERY_DELETES) |
2960 | s += ",recovery_deletes"; | |
181888fb FG |
2961 | if (f & CEPH_OSDMAP_PURGED_SNAPDIRS) |
2962 | s += ",purged_snapdirs"; | |
7c673cae FG |
2963 | if (s.length()) |
2964 | s.erase(0, 1); | |
2965 | return s; | |
2966 | } | |
2967 | ||
2968 | string OSDMap::get_flag_string() const | |
2969 | { | |
2970 | return get_flag_string(flags); | |
2971 | } | |
2972 | ||
7c673cae FG |
2973 | void OSDMap::print_pools(ostream& out) const |
2974 | { | |
2975 | for (const auto &pool : pools) { | |
2976 | std::string name("<unknown>"); | |
2977 | const auto &pni = pool_name.find(pool.first); | |
2978 | if (pni != pool_name.end()) | |
2979 | name = pni->second; | |
2980 | out << "pool " << pool.first | |
2981 | << " '" << name | |
2982 | << "' " << pool.second << "\n"; | |
2983 | ||
2984 | for (const auto &snap : pool.second.snaps) | |
2985 | out << "\tsnap " << snap.second.snapid << " '" << snap.second.name << "' " << snap.second.stamp << "\n"; | |
2986 | ||
2987 | if (!pool.second.removed_snaps.empty()) | |
2988 | out << "\tremoved_snaps " << pool.second.removed_snaps << "\n"; | |
2989 | } | |
2990 | out << std::endl; | |
2991 | } | |
2992 | ||
2993 | void OSDMap::print(ostream& out) const | |
2994 | { | |
2995 | out << "epoch " << get_epoch() << "\n" | |
2996 | << "fsid " << get_fsid() << "\n" | |
2997 | << "created " << get_created() << "\n" | |
2998 | << "modified " << get_modified() << "\n"; | |
2999 | ||
3000 | out << "flags " << get_flag_string() << "\n"; | |
31f18b77 | 3001 | out << "crush_version " << get_crush_version() << "\n"; |
7c673cae FG |
3002 | out << "full_ratio " << full_ratio << "\n"; |
3003 | out << "backfillfull_ratio " << backfillfull_ratio << "\n"; | |
3004 | out << "nearfull_ratio " << nearfull_ratio << "\n"; | |
31f18b77 FG |
3005 | if (require_min_compat_client > 0) { |
3006 | out << "require_min_compat_client " | |
3007 | << ceph_release_name(require_min_compat_client) << "\n"; | |
7c673cae | 3008 | } |
31f18b77 FG |
3009 | out << "min_compat_client " << ceph_release_name(get_min_compat_client()) |
3010 | << "\n"; | |
224ce89b WB |
3011 | if (require_osd_release > 0) { |
3012 | out << "require_osd_release " << ceph_release_name(require_osd_release) | |
3013 | << "\n"; | |
3014 | } | |
7c673cae FG |
3015 | if (get_cluster_snapshot().length()) |
3016 | out << "cluster_snapshot " << get_cluster_snapshot() << "\n"; | |
3017 | out << "\n"; | |
3018 | ||
3019 | print_pools(out); | |
3020 | ||
3021 | out << "max_osd " << get_max_osd() << "\n"; | |
3022 | for (int i=0; i<get_max_osd(); i++) { | |
3023 | if (exists(i)) { | |
3024 | out << "osd." << i; | |
3025 | out << (is_up(i) ? " up ":" down"); | |
3026 | out << (is_in(i) ? " in ":" out"); | |
3027 | out << " weight " << get_weightf(i); | |
3028 | if (get_primary_affinity(i) != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) | |
3029 | out << " primary_affinity " << get_primary_affinityf(i); | |
3030 | const osd_info_t& info(get_info(i)); | |
3031 | out << " " << info; | |
3032 | out << " " << get_addr(i) << " " << get_cluster_addr(i) << " " << get_hb_back_addr(i) | |
3033 | << " " << get_hb_front_addr(i); | |
3034 | set<string> st; | |
3035 | get_state(i, st); | |
3036 | out << " " << st; | |
3037 | if (!get_uuid(i).is_zero()) | |
3038 | out << " " << get_uuid(i); | |
3039 | out << "\n"; | |
3040 | } | |
3041 | } | |
3042 | out << std::endl; | |
3043 | ||
3044 | for (auto& p : pg_upmap) { | |
3045 | out << "pg_upmap " << p.first << " " << p.second << "\n"; | |
3046 | } | |
3047 | for (auto& p : pg_upmap_items) { | |
3048 | out << "pg_upmap_items " << p.first << " " << p.second << "\n"; | |
3049 | } | |
3050 | ||
3051 | for (const auto pg : *pg_temp) | |
3052 | out << "pg_temp " << pg.first << " " << pg.second << "\n"; | |
3053 | ||
3054 | for (const auto pg : *primary_temp) | |
3055 | out << "primary_temp " << pg.first << " " << pg.second << "\n"; | |
3056 | ||
3057 | for (const auto &addr : blacklist) | |
3058 | out << "blacklist " << addr.first << " expires " << addr.second << "\n"; | |
3059 | ||
3060 | // ignore pg_swap_primary | |
3061 | } | |
3062 | ||
3063 | class OSDTreePlainDumper : public CrushTreeDumper::Dumper<TextTable> { | |
3064 | public: | |
3065 | typedef CrushTreeDumper::Dumper<TextTable> Parent; | |
31f18b77 FG |
3066 | |
3067 | OSDTreePlainDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
3068 | unsigned f) | |
c07f9fc5 | 3069 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
3070 | |
3071 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
3072 | if (!filter) { |
3073 | return true; // normal case | |
3074 | } | |
3075 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
3076 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
3077 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
3078 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
3079 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
3080 | return true; | |
31f18b77 | 3081 | } |
c07f9fc5 | 3082 | return false; |
31f18b77 FG |
3083 | } |
3084 | ||
3085 | bool should_dump_empty_bucket() const override { | |
3086 | return !filter; | |
3087 | } | |
7c673cae FG |
3088 | |
3089 | void dump(TextTable *tbl) { | |
3090 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 3091 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
3092 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
3093 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
c07f9fc5 | 3094 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae | 3095 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 3096 | tbl->define_column("PRI-AFF", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
3097 | |
3098 | Parent::dump(tbl); | |
3099 | ||
3100 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
31f18b77 | 3101 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) { |
c07f9fc5 | 3102 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), tbl); |
31f18b77 | 3103 | } |
7c673cae FG |
3104 | } |
3105 | } | |
3106 | ||
3107 | protected: | |
3108 | void dump_item(const CrushTreeDumper::Item &qi, TextTable *tbl) override { | |
224ce89b WB |
3109 | const char *c = crush->get_item_class(qi.id); |
3110 | if (!c) | |
3111 | c = ""; | |
7c673cae | 3112 | *tbl << qi.id |
224ce89b | 3113 | << c |
7c673cae FG |
3114 | << weightf_t(qi.weight); |
3115 | ||
3116 | ostringstream name; | |
3117 | for (int k = 0; k < qi.depth; k++) | |
3118 | name << " "; | |
3119 | if (qi.is_bucket()) { | |
3120 | name << crush->get_type_name(crush->get_bucket_type(qi.id)) << " " | |
3121 | << crush->get_item_name(qi.id); | |
3122 | } else { | |
3123 | name << "osd." << qi.id; | |
3124 | } | |
3125 | *tbl << name.str(); | |
3126 | ||
3127 | if (!qi.is_bucket()) { | |
3128 | if (!osdmap->exists(qi.id)) { | |
3129 | *tbl << "DNE" | |
3130 | << 0; | |
3131 | } else { | |
c07f9fc5 FG |
3132 | string s; |
3133 | if (osdmap->is_up(qi.id)) { | |
3134 | s = "up"; | |
3135 | } else if (osdmap->is_destroyed(qi.id)) { | |
3136 | s = "destroyed"; | |
3137 | } else { | |
3138 | s = "down"; | |
3139 | } | |
3140 | *tbl << s | |
7c673cae FG |
3141 | << weightf_t(osdmap->get_weightf(qi.id)) |
3142 | << weightf_t(osdmap->get_primary_affinityf(qi.id)); | |
3143 | } | |
3144 | } | |
3145 | *tbl << TextTable::endrow; | |
3146 | } | |
3147 | ||
3148 | private: | |
3149 | const OSDMap *osdmap; | |
31f18b77 | 3150 | const unsigned filter; |
7c673cae FG |
3151 | }; |
3152 | ||
3153 | class OSDTreeFormattingDumper : public CrushTreeDumper::FormattingDumper { | |
3154 | public: | |
3155 | typedef CrushTreeDumper::FormattingDumper Parent; | |
3156 | ||
31f18b77 FG |
3157 | OSDTreeFormattingDumper(const CrushWrapper *crush, const OSDMap *osdmap_, |
3158 | unsigned f) | |
c07f9fc5 | 3159 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
3160 | |
3161 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
3162 | if (!filter) { |
3163 | return true; // normal case | |
3164 | } | |
3165 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
3166 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
3167 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
3168 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
3169 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
3170 | return true; | |
31f18b77 | 3171 | } |
c07f9fc5 | 3172 | return false; |
31f18b77 FG |
3173 | } |
3174 | ||
3175 | bool should_dump_empty_bucket() const override { | |
3176 | return !filter; | |
3177 | } | |
7c673cae FG |
3178 | |
3179 | void dump(Formatter *f) { | |
3180 | f->open_array_section("nodes"); | |
3181 | Parent::dump(f); | |
3182 | f->close_section(); | |
3183 | f->open_array_section("stray"); | |
3184 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
31f18b77 | 3185 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) |
c07f9fc5 | 3186 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
7c673cae FG |
3187 | } |
3188 | f->close_section(); | |
3189 | } | |
3190 | ||
3191 | protected: | |
3192 | void dump_item_fields(const CrushTreeDumper::Item &qi, Formatter *f) override { | |
3193 | Parent::dump_item_fields(qi, f); | |
3194 | if (!qi.is_bucket()) | |
3195 | { | |
c07f9fc5 FG |
3196 | string s; |
3197 | if (osdmap->is_up(qi.id)) { | |
3198 | s = "up"; | |
3199 | } else if (osdmap->is_destroyed(qi.id)) { | |
3200 | s = "destroyed"; | |
3201 | } else { | |
3202 | s = "down"; | |
3203 | } | |
7c673cae | 3204 | f->dump_unsigned("exists", (int)osdmap->exists(qi.id)); |
c07f9fc5 | 3205 | f->dump_string("status", s); |
7c673cae FG |
3206 | f->dump_float("reweight", osdmap->get_weightf(qi.id)); |
3207 | f->dump_float("primary_affinity", osdmap->get_primary_affinityf(qi.id)); | |
3208 | } | |
3209 | } | |
3210 | ||
3211 | private: | |
3212 | const OSDMap *osdmap; | |
31f18b77 | 3213 | const unsigned filter; |
7c673cae FG |
3214 | }; |
3215 | ||
31f18b77 | 3216 | void OSDMap::print_tree(Formatter *f, ostream *out, unsigned filter) const |
7c673cae | 3217 | { |
31f18b77 FG |
3218 | if (f) { |
3219 | OSDTreeFormattingDumper(crush.get(), this, filter).dump(f); | |
3220 | } else { | |
7c673cae FG |
3221 | assert(out); |
3222 | TextTable tbl; | |
31f18b77 | 3223 | OSDTreePlainDumper(crush.get(), this, filter).dump(&tbl); |
7c673cae FG |
3224 | *out << tbl; |
3225 | } | |
3226 | } | |
3227 | ||
224ce89b WB |
3228 | void OSDMap::print_summary(Formatter *f, ostream& out, |
3229 | const string& prefix) const | |
7c673cae FG |
3230 | { |
3231 | if (f) { | |
3232 | f->open_object_section("osdmap"); | |
3233 | f->dump_int("epoch", get_epoch()); | |
3234 | f->dump_int("num_osds", get_num_osds()); | |
3235 | f->dump_int("num_up_osds", get_num_up_osds()); | |
3236 | f->dump_int("num_in_osds", get_num_in_osds()); | |
3237 | f->dump_bool("full", test_flag(CEPH_OSDMAP_FULL) ? true : false); | |
3238 | f->dump_bool("nearfull", test_flag(CEPH_OSDMAP_NEARFULL) ? true : false); | |
3239 | f->dump_unsigned("num_remapped_pgs", get_num_pg_temp()); | |
3240 | f->close_section(); | |
3241 | } else { | |
31f18b77 | 3242 | out << get_num_osds() << " osds: " |
7c673cae FG |
3243 | << get_num_up_osds() << " up, " |
3244 | << get_num_in_osds() << " in"; | |
3245 | if (get_num_pg_temp()) | |
3246 | out << "; " << get_num_pg_temp() << " remapped pgs"; | |
3247 | out << "\n"; | |
3248 | uint64_t important_flags = flags & ~CEPH_OSDMAP_SEMIHIDDEN_FLAGS; | |
3249 | if (important_flags) | |
224ce89b | 3250 | out << prefix << "flags " << get_flag_string(important_flags) << "\n"; |
7c673cae FG |
3251 | } |
3252 | } | |
3253 | ||
3254 | void OSDMap::print_oneline_summary(ostream& out) const | |
3255 | { | |
3256 | out << "e" << get_epoch() << ": " | |
31f18b77 | 3257 | << get_num_osds() << " total, " |
7c673cae FG |
3258 | << get_num_up_osds() << " up, " |
3259 | << get_num_in_osds() << " in"; | |
3260 | if (test_flag(CEPH_OSDMAP_FULL)) | |
3261 | out << " full"; | |
3262 | else if (test_flag(CEPH_OSDMAP_NEARFULL)) | |
3263 | out << " nearfull"; | |
3264 | } | |
3265 | ||
3266 | bool OSDMap::crush_ruleset_in_use(int ruleset) const | |
3267 | { | |
3268 | for (const auto &pool : pools) { | |
31f18b77 | 3269 | if (pool.second.crush_rule == ruleset) |
7c673cae FG |
3270 | return true; |
3271 | } | |
3272 | return false; | |
3273 | } | |
3274 | ||
224ce89b WB |
3275 | int OSDMap::build_simple_optioned(CephContext *cct, epoch_t e, uuid_d &fsid, |
3276 | int nosd, int pg_bits, int pgp_bits, | |
3277 | bool default_pool) | |
7c673cae | 3278 | { |
224ce89b WB |
3279 | ldout(cct, 10) << "build_simple on " << nosd |
3280 | << " osds" << dendl; | |
7c673cae FG |
3281 | epoch = e; |
3282 | set_fsid(fsid); | |
3283 | created = modified = ceph_clock_now(); | |
3284 | ||
3285 | if (nosd >= 0) { | |
3286 | set_max_osd(nosd); | |
3287 | } else { | |
3288 | // count osds | |
3289 | int maxosd = 0; | |
3290 | const md_config_t *conf = cct->_conf; | |
3291 | vector<string> sections; | |
3292 | conf->get_all_sections(sections); | |
3293 | ||
3294 | for (auto §ion : sections) { | |
3295 | if (section.find("osd.") != 0) | |
3296 | continue; | |
3297 | ||
3298 | const char *begin = section.c_str() + 4; | |
3299 | char *end = (char*)begin; | |
3300 | int o = strtol(begin, &end, 10); | |
3301 | if (*end != '\0') | |
3302 | continue; | |
3303 | ||
3304 | if (o > cct->_conf->mon_max_osd) { | |
3305 | lderr(cct) << "[osd." << o << "] in config has id > mon_max_osd " << cct->_conf->mon_max_osd << dendl; | |
3306 | return -ERANGE; | |
3307 | } | |
3308 | ||
3309 | if (o > maxosd) | |
3310 | maxosd = o; | |
3311 | } | |
3312 | ||
3313 | set_max_osd(maxosd + 1); | |
3314 | } | |
3315 | ||
7c673cae FG |
3316 | |
3317 | stringstream ss; | |
3318 | int r; | |
3319 | if (nosd >= 0) | |
3320 | r = build_simple_crush_map(cct, *crush, nosd, &ss); | |
3321 | else | |
3322 | r = build_simple_crush_map_from_conf(cct, *crush, &ss); | |
3323 | assert(r == 0); | |
3324 | ||
3325 | int poolbase = get_max_osd() ? get_max_osd() : 1; | |
3326 | ||
d2e6a577 | 3327 | const int default_replicated_rule = crush->get_osd_pool_default_crush_replicated_ruleset(cct); |
31f18b77 | 3328 | assert(default_replicated_rule >= 0); |
7c673cae | 3329 | |
224ce89b WB |
3330 | if (default_pool) { |
3331 | // pgp_num <= pg_num | |
3332 | if (pgp_bits > pg_bits) | |
3333 | pgp_bits = pg_bits; | |
3334 | ||
3335 | vector<string> pool_names; | |
3336 | pool_names.push_back("rbd"); | |
3337 | for (auto &plname : pool_names) { | |
3338 | int64_t pool = ++pool_max; | |
3339 | pools[pool].type = pg_pool_t::TYPE_REPLICATED; | |
3340 | pools[pool].flags = cct->_conf->osd_pool_default_flags; | |
3341 | if (cct->_conf->osd_pool_default_flag_hashpspool) | |
3342 | pools[pool].set_flag(pg_pool_t::FLAG_HASHPSPOOL); | |
3343 | if (cct->_conf->osd_pool_default_flag_nodelete) | |
3344 | pools[pool].set_flag(pg_pool_t::FLAG_NODELETE); | |
3345 | if (cct->_conf->osd_pool_default_flag_nopgchange) | |
3346 | pools[pool].set_flag(pg_pool_t::FLAG_NOPGCHANGE); | |
3347 | if (cct->_conf->osd_pool_default_flag_nosizechange) | |
3348 | pools[pool].set_flag(pg_pool_t::FLAG_NOSIZECHANGE); | |
3349 | pools[pool].size = cct->_conf->osd_pool_default_size; | |
3350 | pools[pool].min_size = cct->_conf->get_osd_pool_default_min_size(); | |
3351 | pools[pool].crush_rule = default_replicated_rule; | |
3352 | pools[pool].object_hash = CEPH_STR_HASH_RJENKINS; | |
3353 | pools[pool].set_pg_num(poolbase << pg_bits); | |
3354 | pools[pool].set_pgp_num(poolbase << pgp_bits); | |
3355 | pools[pool].last_change = epoch; | |
c07f9fc5 FG |
3356 | pools[pool].application_metadata.insert( |
3357 | {pg_pool_t::APPLICATION_NAME_RBD, {}}); | |
224ce89b WB |
3358 | pool_name[pool] = plname; |
3359 | name_pool[plname] = pool; | |
3360 | } | |
7c673cae FG |
3361 | } |
3362 | ||
3363 | for (int i=0; i<get_max_osd(); i++) { | |
3364 | set_state(i, 0); | |
3365 | set_weight(i, CEPH_OSD_OUT); | |
3366 | } | |
3367 | ||
3368 | map<string,string> profile_map; | |
3369 | r = get_erasure_code_profile_default(cct, profile_map, &ss); | |
3370 | if (r < 0) { | |
3371 | lderr(cct) << ss.str() << dendl; | |
3372 | return r; | |
3373 | } | |
3374 | set_erasure_code_profile("default", profile_map); | |
3375 | return 0; | |
3376 | } | |
3377 | ||
3378 | int OSDMap::get_erasure_code_profile_default(CephContext *cct, | |
3379 | map<string,string> &profile_map, | |
3380 | ostream *ss) | |
3381 | { | |
3382 | int r = get_json_str_map(cct->_conf->osd_pool_default_erasure_code_profile, | |
3383 | *ss, | |
3384 | &profile_map); | |
3385 | return r; | |
3386 | } | |
3387 | ||
3388 | int OSDMap::_build_crush_types(CrushWrapper& crush) | |
3389 | { | |
3390 | crush.set_type_name(0, "osd"); | |
3391 | crush.set_type_name(1, "host"); | |
3392 | crush.set_type_name(2, "chassis"); | |
3393 | crush.set_type_name(3, "rack"); | |
3394 | crush.set_type_name(4, "row"); | |
3395 | crush.set_type_name(5, "pdu"); | |
3396 | crush.set_type_name(6, "pod"); | |
3397 | crush.set_type_name(7, "room"); | |
3398 | crush.set_type_name(8, "datacenter"); | |
3399 | crush.set_type_name(9, "region"); | |
3400 | crush.set_type_name(10, "root"); | |
3401 | return 10; | |
3402 | } | |
3403 | ||
3404 | int OSDMap::build_simple_crush_map(CephContext *cct, CrushWrapper& crush, | |
3405 | int nosd, ostream *ss) | |
3406 | { | |
3407 | crush.create(); | |
3408 | ||
3409 | // root | |
3410 | int root_type = _build_crush_types(crush); | |
3411 | int rootid; | |
3412 | int r = crush.add_bucket(0, 0, CRUSH_HASH_DEFAULT, | |
3413 | root_type, 0, NULL, NULL, &rootid); | |
3414 | assert(r == 0); | |
3415 | crush.set_item_name(rootid, "default"); | |
3416 | ||
3417 | for (int o=0; o<nosd; o++) { | |
3418 | map<string,string> loc; | |
3419 | loc["host"] = "localhost"; | |
3420 | loc["rack"] = "localrack"; | |
3421 | loc["root"] = "default"; | |
3422 | ldout(cct, 10) << " adding osd." << o << " at " << loc << dendl; | |
3423 | char name[32]; | |
3424 | snprintf(name, sizeof(name), "osd.%d", o); | |
3425 | crush.insert_item(cct, o, 1.0, name, loc); | |
3426 | } | |
3427 | ||
31f18b77 | 3428 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
3429 | |
3430 | crush.finalize(); | |
3431 | ||
3432 | return 0; | |
3433 | } | |
3434 | ||
3435 | int OSDMap::build_simple_crush_map_from_conf(CephContext *cct, | |
3436 | CrushWrapper& crush, | |
3437 | ostream *ss) | |
3438 | { | |
3439 | const md_config_t *conf = cct->_conf; | |
3440 | ||
3441 | crush.create(); | |
3442 | ||
3443 | // root | |
3444 | int root_type = _build_crush_types(crush); | |
3445 | int rootid; | |
3446 | int r = crush.add_bucket(0, 0, | |
3447 | CRUSH_HASH_DEFAULT, | |
3448 | root_type, 0, NULL, NULL, &rootid); | |
3449 | assert(r == 0); | |
3450 | crush.set_item_name(rootid, "default"); | |
3451 | ||
3452 | // add osds | |
3453 | vector<string> sections; | |
3454 | conf->get_all_sections(sections); | |
3455 | ||
3456 | for (auto §ion : sections) { | |
3457 | if (section.find("osd.") != 0) | |
3458 | continue; | |
3459 | ||
3460 | const char *begin = section.c_str() + 4; | |
3461 | char *end = (char*)begin; | |
3462 | int o = strtol(begin, &end, 10); | |
3463 | if (*end != '\0') | |
3464 | continue; | |
3465 | ||
3466 | string host, rack, row, room, dc, pool; | |
3467 | vector<string> sectiontmp; | |
3468 | sectiontmp.push_back("osd"); | |
3469 | sectiontmp.push_back(section); | |
3470 | conf->get_val_from_conf_file(sectiontmp, "host", host, false); | |
3471 | conf->get_val_from_conf_file(sectiontmp, "rack", rack, false); | |
3472 | conf->get_val_from_conf_file(sectiontmp, "row", row, false); | |
3473 | conf->get_val_from_conf_file(sectiontmp, "room", room, false); | |
3474 | conf->get_val_from_conf_file(sectiontmp, "datacenter", dc, false); | |
3475 | conf->get_val_from_conf_file(sectiontmp, "root", pool, false); | |
3476 | ||
3477 | if (host.length() == 0) | |
3478 | host = "unknownhost"; | |
3479 | if (rack.length() == 0) | |
3480 | rack = "unknownrack"; | |
3481 | ||
3482 | map<string,string> loc; | |
3483 | loc["host"] = host; | |
3484 | loc["rack"] = rack; | |
3485 | if (row.size()) | |
3486 | loc["row"] = row; | |
3487 | if (room.size()) | |
3488 | loc["room"] = room; | |
3489 | if (dc.size()) | |
3490 | loc["datacenter"] = dc; | |
3491 | loc["root"] = "default"; | |
3492 | ||
3493 | ldout(cct, 5) << " adding osd." << o << " at " << loc << dendl; | |
3494 | crush.insert_item(cct, o, 1.0, section, loc); | |
3495 | } | |
3496 | ||
31f18b77 | 3497 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
3498 | |
3499 | crush.finalize(); | |
3500 | ||
3501 | return 0; | |
3502 | } | |
3503 | ||
3504 | ||
31f18b77 FG |
3505 | int OSDMap::build_simple_crush_rules( |
3506 | CephContext *cct, | |
3507 | CrushWrapper& crush, | |
3508 | const string& root, | |
3509 | ostream *ss) | |
7c673cae | 3510 | { |
31f18b77 | 3511 | int crush_rule = crush.get_osd_pool_default_crush_replicated_ruleset(cct); |
7c673cae FG |
3512 | string failure_domain = |
3513 | crush.get_type_name(cct->_conf->osd_crush_chooseleaf_type); | |
3514 | ||
7c673cae | 3515 | int r; |
31f18b77 | 3516 | r = crush.add_simple_rule_at( |
224ce89b | 3517 | "replicated_rule", root, failure_domain, "", |
31f18b77 FG |
3518 | "firstn", pg_pool_t::TYPE_REPLICATED, |
3519 | crush_rule, ss); | |
7c673cae FG |
3520 | if (r < 0) |
3521 | return r; | |
3522 | // do not add an erasure rule by default or else we will implicitly | |
3523 | // require the crush_v2 feature of clients | |
3524 | return 0; | |
3525 | } | |
3526 | ||
3527 | int OSDMap::summarize_mapping_stats( | |
3528 | OSDMap *newmap, | |
3529 | const set<int64_t> *pools, | |
3530 | std::string *out, | |
3531 | Formatter *f) const | |
3532 | { | |
3533 | set<int64_t> ls; | |
3534 | if (pools) { | |
3535 | ls = *pools; | |
3536 | } else { | |
3537 | for (auto &p : get_pools()) | |
3538 | ls.insert(p.first); | |
3539 | } | |
3540 | ||
3541 | unsigned total_pg = 0; | |
3542 | unsigned moved_pg = 0; | |
3543 | vector<unsigned> base_by_osd(get_max_osd(), 0); | |
3544 | vector<unsigned> new_by_osd(get_max_osd(), 0); | |
3545 | for (int64_t pool_id : ls) { | |
3546 | const pg_pool_t *pi = get_pg_pool(pool_id); | |
31f18b77 FG |
3547 | vector<int> up, up2; |
3548 | int up_primary; | |
7c673cae FG |
3549 | for (unsigned ps = 0; ps < pi->get_pg_num(); ++ps) { |
3550 | pg_t pgid(ps, pool_id, -1); | |
3551 | total_pg += pi->get_size(); | |
31f18b77 | 3552 | pg_to_up_acting_osds(pgid, &up, &up_primary, nullptr, nullptr); |
7c673cae FG |
3553 | for (int osd : up) { |
3554 | if (osd >= 0 && osd < get_max_osd()) | |
3555 | ++base_by_osd[osd]; | |
3556 | } | |
3557 | if (newmap) { | |
31f18b77 | 3558 | newmap->pg_to_up_acting_osds(pgid, &up2, &up_primary, nullptr, nullptr); |
7c673cae FG |
3559 | for (int osd : up2) { |
3560 | if (osd >= 0 && osd < get_max_osd()) | |
3561 | ++new_by_osd[osd]; | |
3562 | } | |
3563 | if (pi->type == pg_pool_t::TYPE_ERASURE) { | |
3564 | for (unsigned i=0; i<up.size(); ++i) { | |
3565 | if (up[i] != up2[i]) { | |
3566 | ++moved_pg; | |
3567 | } | |
3568 | } | |
3569 | } else if (pi->type == pg_pool_t::TYPE_REPLICATED) { | |
3570 | for (int osd : up) { | |
3571 | if (std::find(up2.begin(), up2.end(), osd) == up2.end()) { | |
3572 | ++moved_pg; | |
3573 | } | |
3574 | } | |
3575 | } else { | |
3576 | assert(0 == "unhandled pool type"); | |
3577 | } | |
3578 | } | |
3579 | } | |
3580 | } | |
3581 | ||
3582 | unsigned num_up_in = 0; | |
3583 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
3584 | if (is_up(osd) && is_in(osd)) | |
3585 | ++num_up_in; | |
3586 | } | |
3587 | if (!num_up_in) { | |
3588 | return -EINVAL; | |
3589 | } | |
3590 | ||
3591 | float avg_pg = (float)total_pg / (float)num_up_in; | |
3592 | float base_stddev = 0, new_stddev = 0; | |
3593 | int min = -1, max = -1; | |
3594 | unsigned min_base_pg = 0, max_base_pg = 0; | |
3595 | unsigned min_new_pg = 0, max_new_pg = 0; | |
3596 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
3597 | if (is_up(osd) && is_in(osd)) { | |
3598 | float base_diff = (float)base_by_osd[osd] - avg_pg; | |
3599 | base_stddev += base_diff * base_diff; | |
3600 | float new_diff = (float)new_by_osd[osd] - avg_pg; | |
3601 | new_stddev += new_diff * new_diff; | |
3602 | if (min < 0 || base_by_osd[osd] < min_base_pg) { | |
3603 | min = osd; | |
3604 | min_base_pg = base_by_osd[osd]; | |
3605 | min_new_pg = new_by_osd[osd]; | |
3606 | } | |
3607 | if (max < 0 || base_by_osd[osd] > max_base_pg) { | |
3608 | max = osd; | |
3609 | max_base_pg = base_by_osd[osd]; | |
3610 | max_new_pg = new_by_osd[osd]; | |
3611 | } | |
3612 | } | |
3613 | } | |
3614 | base_stddev = sqrt(base_stddev / num_up_in); | |
3615 | new_stddev = sqrt(new_stddev / num_up_in); | |
3616 | ||
3617 | float edev = sqrt(avg_pg * (1.0 - (1.0 / (double)num_up_in))); | |
3618 | ||
3619 | ostringstream ss; | |
3620 | if (f) | |
3621 | f->open_object_section("utilization"); | |
3622 | if (newmap) { | |
3623 | if (f) { | |
3624 | f->dump_unsigned("moved_pgs", moved_pg); | |
3625 | f->dump_unsigned("total_pgs", total_pg); | |
3626 | } else { | |
3627 | float percent = 0; | |
3628 | if (total_pg) | |
3629 | percent = (float)moved_pg * 100.0 / (float)total_pg; | |
3630 | ss << "moved " << moved_pg << " / " << total_pg | |
3631 | << " (" << percent << "%)\n"; | |
3632 | } | |
3633 | } | |
3634 | if (f) { | |
3635 | f->dump_float("avg_pgs", avg_pg); | |
3636 | f->dump_float("std_dev", base_stddev); | |
3637 | f->dump_float("expected_baseline_std_dev", edev); | |
3638 | if (newmap) | |
3639 | f->dump_float("new_std_dev", new_stddev); | |
3640 | } else { | |
3641 | ss << "avg " << avg_pg << "\n"; | |
3642 | ss << "stddev " << base_stddev; | |
3643 | if (newmap) | |
3644 | ss << " -> " << new_stddev; | |
3645 | ss << " (expected baseline " << edev << ")\n"; | |
3646 | } | |
3647 | if (min >= 0) { | |
3648 | if (f) { | |
3649 | f->dump_unsigned("min_osd", min); | |
3650 | f->dump_unsigned("min_osd_pgs", min_base_pg); | |
3651 | if (newmap) | |
3652 | f->dump_unsigned("new_min_osd_pgs", min_new_pg); | |
3653 | } else { | |
3654 | ss << "min osd." << min << " with " << min_base_pg; | |
3655 | if (newmap) | |
3656 | ss << " -> " << min_new_pg; | |
3657 | ss << " pgs (" << (float)min_base_pg / avg_pg; | |
3658 | if (newmap) | |
3659 | ss << " -> " << (float)min_new_pg / avg_pg; | |
3660 | ss << " * mean)\n"; | |
3661 | } | |
3662 | } | |
3663 | if (max >= 0) { | |
3664 | if (f) { | |
3665 | f->dump_unsigned("max_osd", max); | |
3666 | f->dump_unsigned("max_osd_pgs", max_base_pg); | |
3667 | if (newmap) | |
3668 | f->dump_unsigned("new_max_osd_pgs", max_new_pg); | |
3669 | } else { | |
3670 | ss << "max osd." << max << " with " << max_base_pg; | |
3671 | if (newmap) | |
3672 | ss << " -> " << max_new_pg; | |
3673 | ss << " pgs (" << (float)max_base_pg / avg_pg; | |
3674 | if (newmap) | |
3675 | ss << " -> " << (float)max_new_pg / avg_pg; | |
3676 | ss << " * mean)\n"; | |
3677 | } | |
3678 | } | |
3679 | if (f) | |
3680 | f->close_section(); | |
3681 | if (out) | |
3682 | *out = ss.str(); | |
3683 | return 0; | |
3684 | } | |
3685 | ||
3686 | ||
3687 | int OSDMap::clean_pg_upmaps( | |
3688 | CephContext *cct, | |
3689 | Incremental *pending_inc) | |
3690 | { | |
3691 | ldout(cct, 10) << __func__ << dendl; | |
3692 | int changed = 0; | |
3693 | for (auto& p : pg_upmap) { | |
3694 | vector<int> raw; | |
3695 | int primary; | |
3696 | pg_to_raw_osds(p.first, &raw, &primary); | |
3697 | if (vectors_equal(raw, p.second)) { | |
3698 | ldout(cct, 10) << " removing redundant pg_upmap " << p.first << " " | |
3699 | << p.second << dendl; | |
3700 | pending_inc->old_pg_upmap.insert(p.first); | |
3701 | ++changed; | |
3702 | } | |
3703 | } | |
3704 | for (auto& p : pg_upmap_items) { | |
3705 | vector<int> raw; | |
3706 | int primary; | |
3707 | pg_to_raw_osds(p.first, &raw, &primary); | |
3708 | mempool::osdmap::vector<pair<int,int>> newmap; | |
3709 | for (auto& q : p.second) { | |
3710 | if (std::find(raw.begin(), raw.end(), q.first) != raw.end()) { | |
3711 | newmap.push_back(q); | |
3712 | } | |
3713 | } | |
3714 | if (newmap.empty()) { | |
3715 | ldout(cct, 10) << " removing no-op pg_upmap_items " << p.first << " " | |
3716 | << p.second << dendl; | |
3717 | pending_inc->old_pg_upmap_items.insert(p.first); | |
3718 | ++changed; | |
3719 | } else if (newmap != p.second) { | |
3720 | ldout(cct, 10) << " simplifying partially no-op pg_upmap_items " | |
3721 | << p.first << " " << p.second << " -> " << newmap << dendl; | |
3722 | pending_inc->new_pg_upmap_items[p.first] = newmap; | |
3723 | ++changed; | |
3724 | } | |
3725 | } | |
3726 | return changed; | |
3727 | } | |
3728 | ||
3729 | bool OSDMap::try_pg_upmap( | |
3730 | CephContext *cct, | |
3731 | pg_t pg, ///< pg to potentially remap | |
3732 | const set<int>& overfull, ///< osds we'd want to evacuate | |
3733 | const vector<int>& underfull, ///< osds to move to, in order of preference | |
3734 | vector<int> *orig, | |
3735 | vector<int> *out) ///< resulting alternative mapping | |
3736 | { | |
3737 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
3738 | if (!pool) | |
3739 | return false; | |
31f18b77 | 3740 | int rule = crush->find_rule(pool->get_crush_rule(), pool->get_type(), |
7c673cae FG |
3741 | pool->get_size()); |
3742 | if (rule < 0) | |
3743 | return false; | |
3744 | ||
3745 | // get original mapping | |
3746 | _pg_to_raw_osds(*pool, pg, orig, NULL); | |
3747 | ||
3748 | // make sure there is something there to remap | |
3749 | bool any = false; | |
3750 | for (auto osd : *orig) { | |
3751 | if (overfull.count(osd)) { | |
3752 | any = true; | |
3753 | break; | |
3754 | } | |
3755 | } | |
3756 | if (!any) { | |
3757 | return false; | |
3758 | } | |
3759 | ||
3760 | int r = crush->try_remap_rule( | |
3761 | cct, | |
3762 | rule, | |
3763 | pool->get_size(), | |
3764 | overfull, underfull, | |
3765 | *orig, | |
3766 | out); | |
3767 | if (r < 0) | |
3768 | return false; | |
3769 | if (*out == *orig) | |
3770 | return false; | |
3771 | return true; | |
3772 | } | |
3773 | ||
3774 | int OSDMap::calc_pg_upmaps( | |
3775 | CephContext *cct, | |
31f18b77 | 3776 | float max_deviation_ratio, |
7c673cae | 3777 | int max, |
31f18b77 | 3778 | const set<int64_t>& only_pools_orig, |
7c673cae FG |
3779 | OSDMap::Incremental *pending_inc) |
3780 | { | |
31f18b77 FG |
3781 | set<int64_t> only_pools; |
3782 | if (only_pools_orig.empty()) { | |
3783 | for (auto& i : pools) { | |
3784 | only_pools.insert(i.first); | |
3785 | } | |
3786 | } else { | |
3787 | only_pools = only_pools_orig; | |
3788 | } | |
7c673cae FG |
3789 | OSDMap tmp; |
3790 | tmp.deepish_copy_from(*this); | |
31f18b77 FG |
3791 | float start_deviation = 0; |
3792 | float end_deviation = 0; | |
7c673cae FG |
3793 | int num_changed = 0; |
3794 | while (true) { | |
3795 | map<int,set<pg_t>> pgs_by_osd; | |
3796 | int total_pgs = 0; | |
31f18b77 FG |
3797 | float osd_weight_total = 0; |
3798 | map<int,float> osd_weight; | |
7c673cae FG |
3799 | for (auto& i : pools) { |
3800 | if (!only_pools.empty() && !only_pools.count(i.first)) | |
3801 | continue; | |
3802 | for (unsigned ps = 0; ps < i.second.get_pg_num(); ++ps) { | |
3803 | pg_t pg(ps, i.first); | |
3804 | vector<int> up; | |
3805 | tmp.pg_to_up_acting_osds(pg, &up, nullptr, nullptr, nullptr); | |
3806 | for (auto osd : up) { | |
3807 | if (osd != CRUSH_ITEM_NONE) | |
3808 | pgs_by_osd[osd].insert(pg); | |
3809 | } | |
3810 | } | |
3811 | total_pgs += i.second.get_size() * i.second.get_pg_num(); | |
31f18b77 FG |
3812 | |
3813 | map<int,float> pmap; | |
3814 | int ruleno = tmp.crush->find_rule(i.second.get_crush_rule(), | |
3815 | i.second.get_type(), | |
3816 | i.second.get_size()); | |
3817 | tmp.crush->get_rule_weight_osd_map(ruleno, &pmap); | |
3818 | ldout(cct,30) << __func__ << " pool " << i.first << " ruleno " << ruleno << dendl; | |
3819 | for (auto p : pmap) { | |
3820 | osd_weight[p.first] += p.second; | |
3821 | osd_weight_total += p.second; | |
3822 | } | |
7c673cae | 3823 | } |
31f18b77 FG |
3824 | for (auto& i : osd_weight) { |
3825 | int pgs = 0; | |
3826 | auto p = pgs_by_osd.find(i.first); | |
3827 | if (p != pgs_by_osd.end()) | |
3828 | pgs = p->second.size(); | |
3829 | else | |
3830 | pgs_by_osd.emplace(i.first, set<pg_t>()); | |
3831 | ldout(cct, 20) << " osd." << i.first << " weight " << i.second | |
3832 | << " pgs " << pgs << dendl; | |
7c673cae FG |
3833 | } |
3834 | ||
224ce89b WB |
3835 | if (osd_weight_total == 0) { |
3836 | lderr(cct) << __func__ << " abort due to osd_weight_total == 0" << dendl; | |
3837 | break; | |
3838 | } | |
7c673cae FG |
3839 | float pgs_per_weight = total_pgs / osd_weight_total; |
3840 | ldout(cct, 10) << " osd_weight_total " << osd_weight_total << dendl; | |
3841 | ldout(cct, 10) << " pgs_per_weight " << pgs_per_weight << dendl; | |
3842 | ||
3843 | // osd deviation | |
31f18b77 | 3844 | float total_deviation = 0; |
7c673cae FG |
3845 | map<int,float> osd_deviation; // osd, deviation(pgs) |
3846 | multimap<float,int> deviation_osd; // deviation(pgs), osd | |
3847 | set<int> overfull; | |
3848 | for (auto& i : pgs_by_osd) { | |
3849 | float target = osd_weight[i.first] * pgs_per_weight; | |
3850 | float deviation = (float)i.second.size() - target; | |
3851 | ldout(cct, 20) << " osd." << i.first | |
3852 | << "\tpgs " << i.second.size() | |
3853 | << "\ttarget " << target | |
3854 | << "\tdeviation " << deviation | |
3855 | << dendl; | |
3856 | osd_deviation[i.first] = deviation; | |
3857 | deviation_osd.insert(make_pair(deviation, i.first)); | |
31f18b77 | 3858 | if (deviation >= 1.0) |
7c673cae | 3859 | overfull.insert(i.first); |
31f18b77 FG |
3860 | total_deviation += abs(deviation); |
3861 | } | |
3862 | if (num_changed == 0) { | |
3863 | start_deviation = total_deviation; | |
7c673cae | 3864 | } |
31f18b77 | 3865 | end_deviation = total_deviation; |
7c673cae FG |
3866 | |
3867 | // build underfull, sorted from least-full to most-average | |
3868 | vector<int> underfull; | |
3869 | for (auto i = deviation_osd.begin(); | |
3870 | i != deviation_osd.end(); | |
3871 | ++i) { | |
3872 | if (i->first >= -.999) | |
3873 | break; | |
3874 | underfull.push_back(i->second); | |
3875 | } | |
31f18b77 FG |
3876 | ldout(cct, 10) << " total_deviation " << total_deviation |
3877 | << " overfull " << overfull | |
7c673cae FG |
3878 | << " underfull " << underfull << dendl; |
3879 | if (overfull.empty() || underfull.empty()) | |
3880 | break; | |
3881 | ||
3882 | // pick fullest | |
3883 | bool restart = false; | |
3884 | for (auto p = deviation_osd.rbegin(); p != deviation_osd.rend(); ++p) { | |
3885 | int osd = p->second; | |
31f18b77 | 3886 | float deviation = p->first; |
7c673cae | 3887 | float target = osd_weight[osd] * pgs_per_weight; |
224ce89b | 3888 | assert(target > 0); |
31f18b77 | 3889 | if (deviation/target < max_deviation_ratio) { |
7c673cae FG |
3890 | ldout(cct, 10) << " osd." << osd |
3891 | << " target " << target | |
3892 | << " deviation " << deviation | |
31f18b77 FG |
3893 | << " -> ratio " << deviation/target |
3894 | << " < max ratio " << max_deviation_ratio << dendl; | |
7c673cae FG |
3895 | break; |
3896 | } | |
3897 | int num_to_move = deviation; | |
3898 | ldout(cct, 10) << " osd." << osd << " move " << num_to_move << dendl; | |
3899 | if (num_to_move < 1) | |
3900 | break; | |
3901 | ||
3902 | set<pg_t>& pgs = pgs_by_osd[osd]; | |
3903 | ||
3904 | // look for remaps we can un-remap | |
3905 | for (auto pg : pgs) { | |
3906 | auto p = tmp.pg_upmap_items.find(pg); | |
3907 | if (p != tmp.pg_upmap_items.end()) { | |
3908 | for (auto q : p->second) { | |
3909 | if (q.second == osd) { | |
3910 | ldout(cct, 10) << " dropping pg_upmap_items " << pg | |
3911 | << " " << p->second << dendl; | |
3912 | tmp.pg_upmap_items.erase(p); | |
3913 | pending_inc->old_pg_upmap_items.insert(pg); | |
3914 | ++num_changed; | |
3915 | restart = true; | |
3916 | } | |
3917 | } | |
3918 | } | |
3919 | if (restart) | |
3920 | break; | |
3921 | } // pg loop | |
3922 | if (restart) | |
3923 | break; | |
3924 | ||
3925 | for (auto pg : pgs) { | |
3926 | if (tmp.pg_upmap.count(pg) || | |
3927 | tmp.pg_upmap_items.count(pg)) { | |
3928 | ldout(cct, 20) << " already remapped " << pg << dendl; | |
3929 | continue; | |
3930 | } | |
3931 | ldout(cct, 10) << " trying " << pg << dendl; | |
3932 | vector<int> orig, out; | |
3933 | if (!try_pg_upmap(cct, pg, overfull, underfull, &orig, &out)) { | |
3934 | continue; | |
3935 | } | |
3936 | ldout(cct, 10) << " " << pg << " " << orig << " -> " << out << dendl; | |
3937 | if (orig.size() != out.size()) { | |
3938 | continue; | |
3939 | } | |
3940 | assert(orig != out); | |
3941 | auto& rmi = tmp.pg_upmap_items[pg]; | |
3942 | for (unsigned i = 0; i < out.size(); ++i) { | |
3943 | if (orig[i] != out[i]) { | |
3944 | rmi.push_back(make_pair(orig[i], out[i])); | |
3945 | } | |
3946 | } | |
3947 | pending_inc->new_pg_upmap_items[pg] = rmi; | |
3948 | ldout(cct, 10) << " " << pg << " pg_upmap_items " << rmi << dendl; | |
3949 | restart = true; | |
3950 | ++num_changed; | |
3951 | break; | |
3952 | } // pg loop | |
3953 | if (restart) | |
3954 | break; | |
3955 | } // osd loop | |
3956 | ||
3957 | if (!restart) { | |
3958 | ldout(cct, 10) << " failed to find any changes to make" << dendl; | |
3959 | break; | |
3960 | } | |
3961 | if (--max == 0) { | |
3962 | ldout(cct, 10) << " hit max iterations, stopping" << dendl; | |
3963 | break; | |
3964 | } | |
3965 | } | |
31f18b77 FG |
3966 | ldout(cct, 10) << " start deviation " << start_deviation << dendl; |
3967 | ldout(cct, 10) << " end deviation " << end_deviation << dendl; | |
7c673cae FG |
3968 | return num_changed; |
3969 | } | |
31f18b77 FG |
3970 | |
3971 | int OSDMap::get_osds_by_bucket_name(const string &name, set<int> *osds) const | |
3972 | { | |
3973 | return crush->get_leaves(name, osds); | |
3974 | } | |
3975 | ||
3976 | template <typename F> | |
3977 | class OSDUtilizationDumper : public CrushTreeDumper::Dumper<F> { | |
3978 | public: | |
3979 | typedef CrushTreeDumper::Dumper<F> Parent; | |
3980 | ||
3981 | OSDUtilizationDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
3982 | const PGStatService *pgs_, bool tree_) : | |
c07f9fc5 | 3983 | Parent(crush, osdmap_->get_pool_names()), |
31f18b77 FG |
3984 | osdmap(osdmap_), |
3985 | pgs(pgs_), | |
3986 | tree(tree_), | |
3987 | average_util(average_utilization()), | |
3988 | min_var(-1), | |
3989 | max_var(-1), | |
3990 | stddev(0), | |
3991 | sum(0) { | |
3992 | } | |
3993 | ||
3994 | protected: | |
3995 | void dump_stray(F *f) { | |
3996 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
3997 | if (osdmap->exists(i) && !this->is_touched(i)) | |
c07f9fc5 | 3998 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
31f18b77 FG |
3999 | } |
4000 | } | |
4001 | ||
4002 | void dump_item(const CrushTreeDumper::Item &qi, F *f) override { | |
4003 | if (!tree && qi.is_bucket()) | |
4004 | return; | |
4005 | ||
4006 | float reweight = qi.is_bucket() ? -1 : osdmap->get_weightf(qi.id); | |
4007 | int64_t kb = 0, kb_used = 0, kb_avail = 0; | |
4008 | double util = 0; | |
4009 | if (get_bucket_utilization(qi.id, &kb, &kb_used, &kb_avail)) | |
4010 | if (kb_used && kb) | |
4011 | util = 100.0 * (double)kb_used / (double)kb; | |
4012 | ||
4013 | double var = 1.0; | |
4014 | if (average_util) | |
4015 | var = util / average_util; | |
4016 | ||
4017 | size_t num_pgs = qi.is_bucket() ? 0 : pgs->get_num_pg_by_osd(qi.id); | |
4018 | ||
4019 | dump_item(qi, reweight, kb, kb_used, kb_avail, util, var, num_pgs, f); | |
4020 | ||
4021 | if (!qi.is_bucket() && reweight > 0) { | |
4022 | if (min_var < 0 || var < min_var) | |
4023 | min_var = var; | |
4024 | if (max_var < 0 || var > max_var) | |
4025 | max_var = var; | |
4026 | ||
4027 | double dev = util - average_util; | |
4028 | dev *= dev; | |
4029 | stddev += reweight * dev; | |
4030 | sum += reweight; | |
4031 | } | |
4032 | } | |
4033 | ||
4034 | virtual void dump_item(const CrushTreeDumper::Item &qi, | |
4035 | float &reweight, | |
4036 | int64_t kb, | |
4037 | int64_t kb_used, | |
4038 | int64_t kb_avail, | |
4039 | double& util, | |
4040 | double& var, | |
4041 | const size_t num_pgs, | |
4042 | F *f) = 0; | |
4043 | ||
4044 | double dev() { | |
4045 | return sum > 0 ? sqrt(stddev / sum) : 0; | |
4046 | } | |
4047 | ||
4048 | double average_utilization() { | |
4049 | int64_t kb = 0, kb_used = 0; | |
4050 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4051 | if (!osdmap->exists(i) || osdmap->get_weight(i) == 0) | |
4052 | continue; | |
4053 | int64_t kb_i, kb_used_i, kb_avail_i; | |
4054 | if (get_osd_utilization(i, &kb_i, &kb_used_i, &kb_avail_i)) { | |
4055 | kb += kb_i; | |
4056 | kb_used += kb_used_i; | |
4057 | } | |
4058 | } | |
4059 | return kb > 0 ? 100.0 * (double)kb_used / (double)kb : 0; | |
4060 | } | |
4061 | ||
4062 | bool get_osd_utilization(int id, int64_t* kb, int64_t* kb_used, | |
4063 | int64_t* kb_avail) const { | |
4064 | const osd_stat_t *p = pgs->get_osd_stat(id); | |
4065 | if (!p) return false; | |
4066 | *kb = p->kb; | |
4067 | *kb_used = p->kb_used; | |
4068 | *kb_avail = p->kb_avail; | |
4069 | return *kb > 0; | |
4070 | } | |
4071 | ||
4072 | bool get_bucket_utilization(int id, int64_t* kb, int64_t* kb_used, | |
4073 | int64_t* kb_avail) const { | |
4074 | if (id >= 0) { | |
4075 | if (osdmap->is_out(id)) { | |
4076 | *kb = 0; | |
4077 | *kb_used = 0; | |
4078 | *kb_avail = 0; | |
4079 | return true; | |
4080 | } | |
4081 | return get_osd_utilization(id, kb, kb_used, kb_avail); | |
4082 | } | |
4083 | ||
4084 | *kb = 0; | |
4085 | *kb_used = 0; | |
4086 | *kb_avail = 0; | |
4087 | ||
4088 | for (int k = osdmap->crush->get_bucket_size(id) - 1; k >= 0; k--) { | |
4089 | int item = osdmap->crush->get_bucket_item(id, k); | |
4090 | int64_t kb_i = 0, kb_used_i = 0, kb_avail_i = 0; | |
4091 | if (!get_bucket_utilization(item, &kb_i, &kb_used_i, &kb_avail_i)) | |
4092 | return false; | |
4093 | *kb += kb_i; | |
4094 | *kb_used += kb_used_i; | |
4095 | *kb_avail += kb_avail_i; | |
4096 | } | |
4097 | return *kb > 0; | |
4098 | } | |
4099 | ||
4100 | protected: | |
4101 | const OSDMap *osdmap; | |
4102 | const PGStatService *pgs; | |
4103 | bool tree; | |
4104 | double average_util; | |
4105 | double min_var; | |
4106 | double max_var; | |
4107 | double stddev; | |
4108 | double sum; | |
4109 | }; | |
4110 | ||
4111 | ||
4112 | class OSDUtilizationPlainDumper : public OSDUtilizationDumper<TextTable> { | |
4113 | public: | |
4114 | typedef OSDUtilizationDumper<TextTable> Parent; | |
4115 | ||
4116 | OSDUtilizationPlainDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
4117 | const PGStatService *pgs, bool tree) : | |
4118 | Parent(crush, osdmap, pgs, tree) {} | |
4119 | ||
4120 | void dump(TextTable *tbl) { | |
4121 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 4122 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
4123 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
4124 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); | |
4125 | tbl->define_column("SIZE", TextTable::LEFT, TextTable::RIGHT); | |
4126 | tbl->define_column("USE", TextTable::LEFT, TextTable::RIGHT); | |
4127 | tbl->define_column("AVAIL", TextTable::LEFT, TextTable::RIGHT); | |
4128 | tbl->define_column("%USE", TextTable::LEFT, TextTable::RIGHT); | |
4129 | tbl->define_column("VAR", TextTable::LEFT, TextTable::RIGHT); | |
4130 | tbl->define_column("PGS", TextTable::LEFT, TextTable::RIGHT); | |
4131 | if (tree) | |
4132 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
4133 | ||
4134 | Parent::dump(tbl); | |
4135 | ||
4136 | dump_stray(tbl); | |
4137 | ||
224ce89b WB |
4138 | *tbl << "" |
4139 | << "" | |
4140 | << "" << "TOTAL" | |
31f18b77 FG |
4141 | << si_t(pgs->get_osd_sum().kb << 10) |
4142 | << si_t(pgs->get_osd_sum().kb_used << 10) | |
4143 | << si_t(pgs->get_osd_sum().kb_avail << 10) | |
4144 | << lowprecision_t(average_util) | |
4145 | << "" | |
4146 | << TextTable::endrow; | |
4147 | } | |
4148 | ||
4149 | protected: | |
4150 | struct lowprecision_t { | |
4151 | float v; | |
4152 | explicit lowprecision_t(float _v) : v(_v) {} | |
4153 | }; | |
4154 | friend std::ostream &operator<<(ostream& out, const lowprecision_t& v); | |
4155 | ||
4156 | using OSDUtilizationDumper<TextTable>::dump_item; | |
4157 | void dump_item(const CrushTreeDumper::Item &qi, | |
4158 | float &reweight, | |
4159 | int64_t kb, | |
4160 | int64_t kb_used, | |
4161 | int64_t kb_avail, | |
4162 | double& util, | |
4163 | double& var, | |
4164 | const size_t num_pgs, | |
4165 | TextTable *tbl) override { | |
224ce89b WB |
4166 | const char *c = crush->get_item_class(qi.id); |
4167 | if (!c) | |
4168 | c = ""; | |
31f18b77 | 4169 | *tbl << qi.id |
224ce89b | 4170 | << c |
31f18b77 FG |
4171 | << weightf_t(qi.weight) |
4172 | << weightf_t(reweight) | |
4173 | << si_t(kb << 10) | |
4174 | << si_t(kb_used << 10) | |
4175 | << si_t(kb_avail << 10) | |
4176 | << lowprecision_t(util) | |
4177 | << lowprecision_t(var); | |
4178 | ||
4179 | if (qi.is_bucket()) { | |
4180 | *tbl << "-"; | |
4181 | } else { | |
4182 | *tbl << num_pgs; | |
4183 | } | |
4184 | ||
4185 | if (tree) { | |
4186 | ostringstream name; | |
4187 | for (int k = 0; k < qi.depth; k++) | |
4188 | name << " "; | |
4189 | if (qi.is_bucket()) { | |
4190 | int type = crush->get_bucket_type(qi.id); | |
4191 | name << crush->get_type_name(type) << " " | |
4192 | << crush->get_item_name(qi.id); | |
4193 | } else { | |
4194 | name << "osd." << qi.id; | |
4195 | } | |
4196 | *tbl << name.str(); | |
4197 | } | |
4198 | ||
4199 | *tbl << TextTable::endrow; | |
4200 | } | |
4201 | ||
4202 | public: | |
4203 | string summary() { | |
4204 | ostringstream out; | |
4205 | out << "MIN/MAX VAR: " << lowprecision_t(min_var) | |
4206 | << "/" << lowprecision_t(max_var) << " " | |
4207 | << "STDDEV: " << lowprecision_t(dev()); | |
4208 | return out.str(); | |
4209 | } | |
4210 | }; | |
4211 | ||
4212 | ostream& operator<<(ostream& out, | |
4213 | const OSDUtilizationPlainDumper::lowprecision_t& v) | |
4214 | { | |
4215 | if (v.v < -0.01) { | |
4216 | return out << "-"; | |
4217 | } else if (v.v < 0.001) { | |
4218 | return out << "0"; | |
4219 | } else { | |
4220 | std::streamsize p = out.precision(); | |
4221 | return out << std::fixed << std::setprecision(2) << v.v << std::setprecision(p); | |
4222 | } | |
4223 | } | |
4224 | ||
4225 | class OSDUtilizationFormatDumper : public OSDUtilizationDumper<Formatter> { | |
4226 | public: | |
4227 | typedef OSDUtilizationDumper<Formatter> Parent; | |
4228 | ||
4229 | OSDUtilizationFormatDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
4230 | const PGStatService *pgs, bool tree) : | |
4231 | Parent(crush, osdmap, pgs, tree) {} | |
4232 | ||
4233 | void dump(Formatter *f) { | |
4234 | f->open_array_section("nodes"); | |
4235 | Parent::dump(f); | |
4236 | f->close_section(); | |
4237 | ||
4238 | f->open_array_section("stray"); | |
4239 | dump_stray(f); | |
4240 | f->close_section(); | |
4241 | } | |
4242 | ||
4243 | protected: | |
4244 | using OSDUtilizationDumper<Formatter>::dump_item; | |
4245 | void dump_item(const CrushTreeDumper::Item &qi, | |
4246 | float &reweight, | |
4247 | int64_t kb, | |
4248 | int64_t kb_used, | |
4249 | int64_t kb_avail, | |
4250 | double& util, | |
4251 | double& var, | |
4252 | const size_t num_pgs, | |
4253 | Formatter *f) override { | |
4254 | f->open_object_section("item"); | |
c07f9fc5 | 4255 | CrushTreeDumper::dump_item_fields(crush, weight_set_names, qi, f); |
31f18b77 FG |
4256 | f->dump_float("reweight", reweight); |
4257 | f->dump_int("kb", kb); | |
4258 | f->dump_int("kb_used", kb_used); | |
4259 | f->dump_int("kb_avail", kb_avail); | |
4260 | f->dump_float("utilization", util); | |
4261 | f->dump_float("var", var); | |
4262 | f->dump_unsigned("pgs", num_pgs); | |
4263 | CrushTreeDumper::dump_bucket_children(crush, qi, f); | |
4264 | f->close_section(); | |
4265 | } | |
4266 | ||
4267 | public: | |
4268 | void summary(Formatter *f) { | |
4269 | f->open_object_section("summary"); | |
4270 | f->dump_int("total_kb", pgs->get_osd_sum().kb); | |
4271 | f->dump_int("total_kb_used", pgs->get_osd_sum().kb_used); | |
4272 | f->dump_int("total_kb_avail", pgs->get_osd_sum().kb_avail); | |
4273 | f->dump_float("average_utilization", average_util); | |
4274 | f->dump_float("min_var", min_var); | |
4275 | f->dump_float("max_var", max_var); | |
4276 | f->dump_float("dev", dev()); | |
4277 | f->close_section(); | |
4278 | } | |
4279 | }; | |
4280 | ||
4281 | void print_osd_utilization(const OSDMap& osdmap, | |
4282 | const PGStatService *pgstat, | |
4283 | ostream& out, | |
4284 | Formatter *f, | |
4285 | bool tree) | |
4286 | { | |
4287 | const CrushWrapper *crush = osdmap.crush.get(); | |
4288 | if (f) { | |
4289 | f->open_object_section("df"); | |
4290 | OSDUtilizationFormatDumper d(crush, &osdmap, pgstat, tree); | |
4291 | d.dump(f); | |
4292 | d.summary(f); | |
4293 | f->close_section(); | |
4294 | f->flush(out); | |
4295 | } else { | |
4296 | OSDUtilizationPlainDumper d(crush, &osdmap, pgstat, tree); | |
4297 | TextTable tbl; | |
4298 | d.dump(&tbl); | |
4299 | out << tbl << d.summary() << "\n"; | |
4300 | } | |
4301 | } | |
224ce89b WB |
4302 | |
4303 | void OSDMap::check_health(health_check_map_t *checks) const | |
4304 | { | |
4305 | int num_osds = get_num_osds(); | |
4306 | ||
4307 | // OSD_DOWN | |
4308 | // OSD_$subtree_DOWN | |
4309 | // OSD_ORPHAN | |
4310 | if (num_osds >= 0) { | |
4311 | int num_in_osds = 0; | |
4312 | int num_down_in_osds = 0; | |
4313 | set<int> osds; | |
4314 | set<int> down_in_osds; | |
4315 | set<int> up_in_osds; | |
4316 | set<int> subtree_up; | |
4317 | unordered_map<int, set<int> > subtree_type_down; | |
4318 | unordered_map<int, int> num_osds_subtree; | |
4319 | int max_type = crush->get_max_type_id(); | |
4320 | ||
4321 | for (int i = 0; i < get_max_osd(); i++) { | |
4322 | if (!exists(i)) { | |
4323 | if (crush->item_exists(i)) { | |
4324 | osds.insert(i); | |
4325 | } | |
4326 | continue; | |
4327 | } | |
4328 | if (is_out(i)) | |
4329 | continue; | |
4330 | ++num_in_osds; | |
4331 | if (down_in_osds.count(i) || up_in_osds.count(i)) | |
4332 | continue; | |
4333 | if (!is_up(i)) { | |
4334 | down_in_osds.insert(i); | |
4335 | int parent_id = 0; | |
4336 | int current = i; | |
4337 | for (int type = 0; type <= max_type; type++) { | |
4338 | if (!crush->get_type_name(type)) | |
4339 | continue; | |
4340 | int r = crush->get_immediate_parent_id(current, &parent_id); | |
4341 | if (r == -ENOENT) | |
4342 | break; | |
4343 | // break early if this parent is already marked as up | |
4344 | if (subtree_up.count(parent_id)) | |
4345 | break; | |
4346 | type = crush->get_bucket_type(parent_id); | |
4347 | if (!subtree_type_is_down( | |
4348 | g_ceph_context, parent_id, type, | |
4349 | &down_in_osds, &up_in_osds, &subtree_up, &subtree_type_down)) | |
4350 | break; | |
4351 | current = parent_id; | |
4352 | } | |
4353 | } | |
4354 | } | |
4355 | ||
4356 | // calculate the number of down osds in each down subtree and | |
4357 | // store it in num_osds_subtree | |
4358 | for (int type = 1; type <= max_type; type++) { | |
4359 | if (!crush->get_type_name(type)) | |
4360 | continue; | |
4361 | for (auto j = subtree_type_down[type].begin(); | |
4362 | j != subtree_type_down[type].end(); | |
4363 | ++j) { | |
4364 | list<int> children; | |
4365 | int num = 0; | |
4366 | int num_children = crush->get_children(*j, &children); | |
4367 | if (num_children == 0) | |
4368 | continue; | |
4369 | for (auto l = children.begin(); l != children.end(); ++l) { | |
4370 | if (*l >= 0) { | |
4371 | ++num; | |
4372 | } else if (num_osds_subtree[*l] > 0) { | |
4373 | num = num + num_osds_subtree[*l]; | |
4374 | } | |
4375 | } | |
4376 | num_osds_subtree[*j] = num; | |
4377 | } | |
4378 | } | |
4379 | num_down_in_osds = down_in_osds.size(); | |
4380 | assert(num_down_in_osds <= num_in_osds); | |
4381 | if (num_down_in_osds > 0) { | |
4382 | // summary of down subtree types and osds | |
4383 | for (int type = max_type; type > 0; type--) { | |
4384 | if (!crush->get_type_name(type)) | |
4385 | continue; | |
4386 | if (subtree_type_down[type].size() > 0) { | |
4387 | ostringstream ss; | |
4388 | ss << subtree_type_down[type].size() << " " | |
4389 | << crush->get_type_name(type); | |
4390 | if (subtree_type_down[type].size() > 1) { | |
4391 | ss << "s"; | |
4392 | } | |
4393 | int sum_down_osds = 0; | |
4394 | for (auto j = subtree_type_down[type].begin(); | |
4395 | j != subtree_type_down[type].end(); | |
4396 | ++j) { | |
4397 | sum_down_osds = sum_down_osds + num_osds_subtree[*j]; | |
4398 | } | |
4399 | ss << " (" << sum_down_osds << " osds) down"; | |
4400 | string err = string("OSD_") + | |
4401 | string(crush->get_type_name(type)) + "_DOWN"; | |
4402 | boost::to_upper(err); | |
4403 | auto& d = checks->add(err, HEALTH_WARN, ss.str()); | |
4404 | for (auto j = subtree_type_down[type].rbegin(); | |
4405 | j != subtree_type_down[type].rend(); | |
4406 | ++j) { | |
4407 | ostringstream ss; | |
4408 | ss << crush->get_type_name(type); | |
4409 | ss << " "; | |
4410 | ss << crush->get_item_name(*j); | |
4411 | // at the top level, do not print location | |
4412 | if (type != max_type) { | |
4413 | ss << " ("; | |
4414 | ss << crush->get_full_location_ordered_string(*j); | |
4415 | ss << ")"; | |
4416 | } | |
4417 | int num = num_osds_subtree[*j]; | |
4418 | ss << " (" << num << " osds)"; | |
4419 | ss << " is down"; | |
4420 | d.detail.push_back(ss.str()); | |
4421 | } | |
4422 | } | |
4423 | } | |
4424 | ostringstream ss; | |
4425 | ss << down_in_osds.size() << " osds down"; | |
4426 | auto& d = checks->add("OSD_DOWN", HEALTH_WARN, ss.str()); | |
4427 | for (auto it = down_in_osds.begin(); it != down_in_osds.end(); ++it) { | |
4428 | ostringstream ss; | |
4429 | ss << "osd." << *it << " ("; | |
4430 | ss << crush->get_full_location_ordered_string(*it); | |
4431 | ss << ") is down"; | |
4432 | d.detail.push_back(ss.str()); | |
4433 | } | |
4434 | } | |
4435 | ||
4436 | if (!osds.empty()) { | |
4437 | ostringstream ss; | |
4438 | ss << osds.size() << " osds exist in the crush map but not in the osdmap"; | |
4439 | auto& d = checks->add("OSD_ORPHAN", HEALTH_WARN, ss.str()); | |
4440 | for (auto osd : osds) { | |
4441 | ostringstream ss; | |
4442 | ss << "osd." << osd << " exists in crush map but not in osdmap"; | |
4443 | d.detail.push_back(ss.str()); | |
4444 | } | |
4445 | } | |
4446 | } | |
4447 | ||
4448 | // OSD_OUT_OF_ORDER_FULL | |
4449 | { | |
4450 | // An osd could configure failsafe ratio, to something different | |
4451 | // but for now assume it is the same here. | |
4452 | float fsr = g_conf->osd_failsafe_full_ratio; | |
4453 | if (fsr > 1.0) fsr /= 100; | |
4454 | float fr = get_full_ratio(); | |
4455 | float br = get_backfillfull_ratio(); | |
4456 | float nr = get_nearfull_ratio(); | |
4457 | ||
4458 | list<string> detail; | |
4459 | // These checks correspond to how OSDService::check_full_status() in an OSD | |
4460 | // handles the improper setting of these values. | |
4461 | if (br < nr) { | |
4462 | ostringstream ss; | |
4463 | ss << "backfillfull_ratio (" << br | |
4464 | << ") < nearfull_ratio (" << nr << "), increased"; | |
4465 | detail.push_back(ss.str()); | |
4466 | br = nr; | |
4467 | } | |
4468 | if (fr < br) { | |
4469 | ostringstream ss; | |
4470 | ss << "full_ratio (" << fr << ") < backfillfull_ratio (" << br | |
4471 | << "), increased"; | |
4472 | detail.push_back(ss.str()); | |
4473 | fr = br; | |
4474 | } | |
4475 | if (fsr < fr) { | |
4476 | ostringstream ss; | |
4477 | ss << "osd_failsafe_full_ratio (" << fsr << ") < full_ratio (" << fr | |
4478 | << "), increased"; | |
4479 | detail.push_back(ss.str()); | |
4480 | } | |
4481 | if (!detail.empty()) { | |
4482 | auto& d = checks->add("OSD_OUT_OF_ORDER_FULL", HEALTH_ERR, | |
4483 | "full ratio(s) out of order"); | |
4484 | d.detail.swap(detail); | |
4485 | } | |
4486 | } | |
4487 | ||
4488 | // OSD_FULL | |
4489 | // OSD_NEARFULL | |
4490 | // OSD_BACKFILLFULL | |
4491 | // OSD_FAILSAFE_FULL | |
4492 | { | |
4493 | set<int> full, backfillfull, nearfull; | |
4494 | get_full_osd_counts(&full, &backfillfull, &nearfull); | |
4495 | if (full.size()) { | |
4496 | ostringstream ss; | |
4497 | ss << full.size() << " full osd(s)"; | |
4498 | auto& d = checks->add("OSD_FULL", HEALTH_ERR, ss.str()); | |
4499 | for (auto& i: full) { | |
4500 | ostringstream ss; | |
4501 | ss << "osd." << i << " is full"; | |
4502 | d.detail.push_back(ss.str()); | |
4503 | } | |
4504 | } | |
4505 | if (backfillfull.size()) { | |
4506 | ostringstream ss; | |
4507 | ss << backfillfull.size() << " backfillfull osd(s)"; | |
4508 | auto& d = checks->add("OSD_BACKFILLFULL", HEALTH_WARN, ss.str()); | |
4509 | for (auto& i: backfillfull) { | |
4510 | ostringstream ss; | |
4511 | ss << "osd." << i << " is backfill full"; | |
4512 | d.detail.push_back(ss.str()); | |
4513 | } | |
4514 | } | |
4515 | if (nearfull.size()) { | |
4516 | ostringstream ss; | |
4517 | ss << nearfull.size() << " nearfull osd(s)"; | |
4518 | auto& d = checks->add("OSD_NEARFULL", HEALTH_WARN, ss.str()); | |
4519 | for (auto& i: nearfull) { | |
4520 | ostringstream ss; | |
4521 | ss << "osd." << i << " is near full"; | |
4522 | d.detail.push_back(ss.str()); | |
4523 | } | |
4524 | } | |
4525 | } | |
4526 | ||
4527 | // OSDMAP_FLAGS | |
4528 | { | |
4529 | // warn about flags | |
4530 | uint64_t warn_flags = | |
4531 | CEPH_OSDMAP_FULL | | |
4532 | CEPH_OSDMAP_PAUSERD | | |
4533 | CEPH_OSDMAP_PAUSEWR | | |
4534 | CEPH_OSDMAP_PAUSEREC | | |
4535 | CEPH_OSDMAP_NOUP | | |
4536 | CEPH_OSDMAP_NODOWN | | |
4537 | CEPH_OSDMAP_NOIN | | |
4538 | CEPH_OSDMAP_NOOUT | | |
4539 | CEPH_OSDMAP_NOBACKFILL | | |
4540 | CEPH_OSDMAP_NORECOVER | | |
4541 | CEPH_OSDMAP_NOSCRUB | | |
4542 | CEPH_OSDMAP_NODEEP_SCRUB | | |
4543 | CEPH_OSDMAP_NOTIERAGENT | | |
4544 | CEPH_OSDMAP_NOREBALANCE; | |
4545 | if (test_flag(warn_flags)) { | |
4546 | ostringstream ss; | |
4547 | ss << get_flag_string(get_flags() & warn_flags) | |
4548 | << " flag(s) set"; | |
4549 | checks->add("OSDMAP_FLAGS", HEALTH_WARN, ss.str()); | |
4550 | } | |
4551 | } | |
4552 | ||
4553 | // OSD_FLAGS | |
4554 | { | |
4555 | list<string> detail; | |
4556 | const unsigned flags = | |
4557 | CEPH_OSD_NOUP | | |
4558 | CEPH_OSD_NOIN | | |
4559 | CEPH_OSD_NODOWN | | |
4560 | CEPH_OSD_NOOUT; | |
4561 | for (int i = 0; i < max_osd; ++i) { | |
4562 | if (osd_state[i] & flags) { | |
4563 | ostringstream ss; | |
4564 | set<string> states; | |
4565 | OSDMap::calc_state_set(osd_state[i] & flags, states); | |
4566 | ss << "osd." << i << " has flags " << states; | |
4567 | detail.push_back(ss.str()); | |
4568 | } | |
4569 | } | |
4570 | if (!detail.empty()) { | |
4571 | ostringstream ss; | |
4572 | ss << detail.size() << " osd(s) have {NOUP,NODOWN,NOIN,NOOUT} flags set"; | |
4573 | auto& d = checks->add("OSD_FLAGS", HEALTH_WARN, ss.str()); | |
4574 | d.detail.swap(detail); | |
4575 | } | |
4576 | } | |
4577 | ||
4578 | // OLD_CRUSH_TUNABLES | |
4579 | if (g_conf->mon_warn_on_legacy_crush_tunables) { | |
4580 | string min = crush->get_min_required_version(); | |
4581 | if (min < g_conf->mon_crush_min_required_version) { | |
4582 | ostringstream ss; | |
4583 | ss << "crush map has legacy tunables (require " << min | |
4584 | << ", min is " << g_conf->mon_crush_min_required_version << ")"; | |
4585 | auto& d = checks->add("OLD_CRUSH_TUNABLES", HEALTH_WARN, ss.str()); | |
4586 | d.detail.push_back("see http://docs.ceph.com/docs/master/rados/operations/crush-map/#tunables"); | |
4587 | } | |
4588 | } | |
4589 | ||
4590 | // OLD_CRUSH_STRAW_CALC_VERSION | |
4591 | if (g_conf->mon_warn_on_crush_straw_calc_version_zero) { | |
4592 | if (crush->get_straw_calc_version() == 0) { | |
4593 | ostringstream ss; | |
4594 | ss << "crush map has straw_calc_version=0"; | |
4595 | auto& d = checks->add("OLD_CRUSH_STRAW_CALC_VERSION", HEALTH_WARN, ss.str()); | |
4596 | d.detail.push_back( | |
4597 | "see http://docs.ceph.com/docs/master/rados/operations/crush-map/#tunables"); | |
4598 | } | |
4599 | } | |
4600 | ||
4601 | // CACHE_POOL_NO_HIT_SET | |
4602 | if (g_conf->mon_warn_on_cache_pools_without_hit_sets) { | |
4603 | list<string> detail; | |
4604 | for (map<int64_t, pg_pool_t>::const_iterator p = pools.begin(); | |
4605 | p != pools.end(); | |
4606 | ++p) { | |
4607 | const pg_pool_t& info = p->second; | |
4608 | if (info.cache_mode_requires_hit_set() && | |
4609 | info.hit_set_params.get_type() == HitSet::TYPE_NONE) { | |
4610 | ostringstream ss; | |
4611 | ss << "pool '" << get_pool_name(p->first) | |
4612 | << "' with cache_mode " << info.get_cache_mode_name() | |
4613 | << " needs hit_set_type to be set but it is not"; | |
4614 | detail.push_back(ss.str()); | |
4615 | } | |
4616 | } | |
4617 | if (!detail.empty()) { | |
4618 | ostringstream ss; | |
4619 | ss << detail.size() << " cache pools are missing hit_sets"; | |
4620 | auto& d = checks->add("CACHE_POOL_NO_HIT_SET", HEALTH_WARN, ss.str()); | |
4621 | d.detail.swap(detail); | |
4622 | } | |
4623 | } | |
4624 | ||
4625 | // OSD_NO_SORTBITWISE | |
4626 | if (!test_flag(CEPH_OSDMAP_SORTBITWISE) && | |
4627 | (get_up_osd_features() & | |
4628 | CEPH_FEATURE_OSD_BITWISE_HOBJ_SORT)) { | |
4629 | ostringstream ss; | |
4630 | ss << "no legacy OSD present but 'sortbitwise' flag is not set"; | |
4631 | checks->add("OSD_NO_SORTBITWISE", HEALTH_WARN, ss.str()); | |
4632 | } | |
4633 | ||
4634 | // OSD_UPGRADE_FINISHED | |
4635 | // none of these (yet) since we don't run until luminous upgrade is done. | |
4636 | ||
4637 | // POOL_FULL | |
4638 | { | |
4639 | list<string> detail; | |
4640 | for (auto it : get_pools()) { | |
4641 | const pg_pool_t &pool = it.second; | |
4642 | if (pool.has_flag(pg_pool_t::FLAG_FULL)) { | |
4643 | const string& pool_name = get_pool_name(it.first); | |
4644 | stringstream ss; | |
4645 | ss << "pool '" << pool_name << "' is full"; | |
4646 | detail.push_back(ss.str()); | |
4647 | } | |
4648 | } | |
4649 | if (!detail.empty()) { | |
4650 | ostringstream ss; | |
4651 | ss << detail.size() << " pool(s) full"; | |
4652 | auto& d = checks->add("POOL_FULL", HEALTH_WARN, ss.str()); | |
4653 | d.detail.swap(detail); | |
4654 | } | |
4655 | } | |
4656 | } | |
35e4c445 FG |
4657 | |
4658 | int OSDMap::parse_osd_id_list(const vector<string>& ls, set<int> *out, | |
4659 | ostream *ss) const | |
4660 | { | |
4661 | out->clear(); | |
4662 | for (auto i = ls.begin(); i != ls.end(); ++i) { | |
4663 | if (i == ls.begin() && | |
4664 | (*i == "any" || *i == "all" || *i == "*")) { | |
4665 | get_all_osds(*out); | |
4666 | break; | |
4667 | } | |
4668 | long osd = parse_osd_id(i->c_str(), ss); | |
4669 | if (osd < 0) { | |
4670 | *ss << "invalid osd id '" << *i << "'"; | |
4671 | return -EINVAL; | |
4672 | } | |
4673 | out->insert(osd); | |
4674 | } | |
4675 | return 0; | |
4676 | } |