]>
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 | ||
11fdf7f2 TL |
18 | #include <algorithm> |
19 | #include <optional> | |
20 | #include <random> | |
21 | ||
224ce89b WB |
22 | #include <boost/algorithm/string.hpp> |
23 | ||
7c673cae | 24 | #include "OSDMap.h" |
7c673cae | 25 | #include "common/config.h" |
3efd9988 | 26 | #include "common/errno.h" |
7c673cae FG |
27 | #include "common/Formatter.h" |
28 | #include "common/TextTable.h" | |
29 | #include "include/ceph_features.h" | |
30 | #include "include/str_map.h" | |
31 | ||
32 | #include "common/code_environment.h" | |
224ce89b | 33 | #include "mon/health_check.h" |
7c673cae FG |
34 | |
35 | #include "crush/CrushTreeDumper.h" | |
36 | #include "common/Clock.h" | |
11fdf7f2 TL |
37 | #include "mon/PGMap.h" |
38 | ||
7c673cae FG |
39 | #define dout_subsys ceph_subsys_osd |
40 | ||
41 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap, osdmap, osdmap); | |
42 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap::Incremental, osdmap_inc, osdmap); | |
43 | ||
44 | ||
45 | // ---------------------------------- | |
46 | // osd_info_t | |
47 | ||
48 | void osd_info_t::dump(Formatter *f) const | |
49 | { | |
50 | f->dump_int("last_clean_begin", last_clean_begin); | |
51 | f->dump_int("last_clean_end", last_clean_end); | |
52 | f->dump_int("up_from", up_from); | |
53 | f->dump_int("up_thru", up_thru); | |
54 | f->dump_int("down_at", down_at); | |
55 | f->dump_int("lost_at", lost_at); | |
56 | } | |
57 | ||
58 | void osd_info_t::encode(bufferlist& bl) const | |
59 | { | |
11fdf7f2 | 60 | using ceph::encode; |
7c673cae | 61 | __u8 struct_v = 1; |
11fdf7f2 TL |
62 | encode(struct_v, bl); |
63 | encode(last_clean_begin, bl); | |
64 | encode(last_clean_end, bl); | |
65 | encode(up_from, bl); | |
66 | encode(up_thru, bl); | |
67 | encode(down_at, bl); | |
68 | encode(lost_at, bl); | |
7c673cae FG |
69 | } |
70 | ||
11fdf7f2 | 71 | void osd_info_t::decode(bufferlist::const_iterator& bl) |
7c673cae | 72 | { |
11fdf7f2 | 73 | using ceph::decode; |
7c673cae | 74 | __u8 struct_v; |
11fdf7f2 TL |
75 | decode(struct_v, bl); |
76 | decode(last_clean_begin, bl); | |
77 | decode(last_clean_end, bl); | |
78 | decode(up_from, bl); | |
79 | decode(up_thru, bl); | |
80 | decode(down_at, bl); | |
81 | decode(lost_at, bl); | |
7c673cae FG |
82 | } |
83 | ||
84 | void osd_info_t::generate_test_instances(list<osd_info_t*>& o) | |
85 | { | |
86 | o.push_back(new osd_info_t); | |
87 | o.push_back(new osd_info_t); | |
88 | o.back()->last_clean_begin = 1; | |
89 | o.back()->last_clean_end = 2; | |
90 | o.back()->up_from = 30; | |
91 | o.back()->up_thru = 40; | |
92 | o.back()->down_at = 5; | |
93 | o.back()->lost_at = 6; | |
94 | } | |
95 | ||
96 | ostream& operator<<(ostream& out, const osd_info_t& info) | |
97 | { | |
98 | out << "up_from " << info.up_from | |
99 | << " up_thru " << info.up_thru | |
100 | << " down_at " << info.down_at | |
101 | << " last_clean_interval [" << info.last_clean_begin << "," << info.last_clean_end << ")"; | |
102 | if (info.lost_at) | |
103 | out << " lost_at " << info.lost_at; | |
104 | return out; | |
105 | } | |
106 | ||
107 | // ---------------------------------- | |
108 | // osd_xinfo_t | |
109 | ||
110 | void osd_xinfo_t::dump(Formatter *f) const | |
111 | { | |
112 | f->dump_stream("down_stamp") << down_stamp; | |
113 | f->dump_float("laggy_probability", laggy_probability); | |
114 | f->dump_int("laggy_interval", laggy_interval); | |
115 | f->dump_int("features", features); | |
116 | f->dump_unsigned("old_weight", old_weight); | |
117 | } | |
118 | ||
119 | void osd_xinfo_t::encode(bufferlist& bl) const | |
120 | { | |
121 | ENCODE_START(3, 1, bl); | |
11fdf7f2 | 122 | encode(down_stamp, bl); |
7c673cae | 123 | __u32 lp = laggy_probability * 0xfffffffful; |
11fdf7f2 TL |
124 | encode(lp, bl); |
125 | encode(laggy_interval, bl); | |
126 | encode(features, bl); | |
127 | encode(old_weight, bl); | |
7c673cae FG |
128 | ENCODE_FINISH(bl); |
129 | } | |
130 | ||
11fdf7f2 | 131 | void osd_xinfo_t::decode(bufferlist::const_iterator& bl) |
7c673cae FG |
132 | { |
133 | DECODE_START(3, bl); | |
11fdf7f2 | 134 | decode(down_stamp, bl); |
7c673cae | 135 | __u32 lp; |
11fdf7f2 | 136 | decode(lp, bl); |
7c673cae | 137 | laggy_probability = (float)lp / (float)0xffffffff; |
11fdf7f2 | 138 | decode(laggy_interval, bl); |
7c673cae | 139 | if (struct_v >= 2) |
11fdf7f2 | 140 | decode(features, bl); |
7c673cae FG |
141 | else |
142 | features = 0; | |
143 | if (struct_v >= 3) | |
11fdf7f2 | 144 | decode(old_weight, bl); |
7c673cae FG |
145 | else |
146 | old_weight = 0; | |
147 | DECODE_FINISH(bl); | |
148 | } | |
149 | ||
150 | void osd_xinfo_t::generate_test_instances(list<osd_xinfo_t*>& o) | |
151 | { | |
152 | o.push_back(new osd_xinfo_t); | |
153 | o.push_back(new osd_xinfo_t); | |
154 | o.back()->down_stamp = utime_t(2, 3); | |
155 | o.back()->laggy_probability = .123; | |
156 | o.back()->laggy_interval = 123456; | |
157 | o.back()->old_weight = 0x7fff; | |
158 | } | |
159 | ||
160 | ostream& operator<<(ostream& out, const osd_xinfo_t& xi) | |
161 | { | |
162 | return out << "down_stamp " << xi.down_stamp | |
163 | << " laggy_probability " << xi.laggy_probability | |
164 | << " laggy_interval " << xi.laggy_interval | |
165 | << " old_weight " << xi.old_weight; | |
166 | } | |
167 | ||
168 | // ---------------------------------- | |
169 | // OSDMap::Incremental | |
170 | ||
171 | int OSDMap::Incremental::get_net_marked_out(const OSDMap *previous) const | |
172 | { | |
173 | int n = 0; | |
174 | for (auto &weight : new_weight) { | |
175 | if (weight.second == CEPH_OSD_OUT && !previous->is_out(weight.first)) | |
176 | n++; // marked out | |
177 | else if (weight.second != CEPH_OSD_OUT && previous->is_out(weight.first)) | |
178 | n--; // marked in | |
179 | } | |
180 | return n; | |
181 | } | |
182 | ||
183 | int OSDMap::Incremental::get_net_marked_down(const OSDMap *previous) const | |
184 | { | |
185 | int n = 0; | |
186 | for (auto &state : new_state) { // | |
187 | if (state.second & CEPH_OSD_UP) { | |
188 | if (previous->is_up(state.first)) | |
189 | n++; // marked down | |
190 | else | |
191 | n--; // marked up | |
192 | } | |
193 | } | |
194 | return n; | |
195 | } | |
196 | ||
197 | int OSDMap::Incremental::identify_osd(uuid_d u) const | |
198 | { | |
199 | for (auto &uuid : new_uuid) | |
200 | if (uuid.second == u) | |
201 | return uuid.first; | |
202 | return -1; | |
203 | } | |
204 | ||
205 | int OSDMap::Incremental::propagate_snaps_to_tiers(CephContext *cct, | |
206 | const OSDMap& osdmap) | |
207 | { | |
11fdf7f2 | 208 | ceph_assert(epoch == osdmap.get_epoch() + 1); |
7c673cae FG |
209 | |
210 | for (auto &new_pool : new_pools) { | |
211 | if (!new_pool.second.tiers.empty()) { | |
212 | pg_pool_t& base = new_pool.second; | |
213 | ||
11fdf7f2 TL |
214 | auto new_rem_it = new_removed_snaps.find(new_pool.first); |
215 | ||
7c673cae FG |
216 | for (const auto &tier_pool : base.tiers) { |
217 | const auto &r = new_pools.find(tier_pool); | |
218 | pg_pool_t *tier = 0; | |
219 | if (r == new_pools.end()) { | |
220 | const pg_pool_t *orig = osdmap.get_pg_pool(tier_pool); | |
221 | if (!orig) { | |
222 | lderr(cct) << __func__ << " no pool " << tier_pool << dendl; | |
223 | return -EIO; | |
224 | } | |
225 | tier = get_new_pool(tier_pool, orig); | |
226 | } else { | |
227 | tier = &r->second; | |
228 | } | |
229 | if (tier->tier_of != new_pool.first) { | |
230 | lderr(cct) << __func__ << " " << r->first << " tier_of != " << new_pool.first << dendl; | |
231 | return -EIO; | |
232 | } | |
233 | ||
234 | ldout(cct, 10) << __func__ << " from " << new_pool.first << " to " | |
235 | << tier_pool << dendl; | |
236 | tier->snap_seq = base.snap_seq; | |
237 | tier->snap_epoch = base.snap_epoch; | |
238 | tier->snaps = base.snaps; | |
239 | tier->removed_snaps = base.removed_snaps; | |
11fdf7f2 TL |
240 | tier->flags |= base.flags & (pg_pool_t::FLAG_SELFMANAGED_SNAPS| |
241 | pg_pool_t::FLAG_POOL_SNAPS); | |
242 | ||
243 | if (new_rem_it != new_removed_snaps.end()) { | |
244 | new_removed_snaps[tier_pool] = new_rem_it->second; | |
245 | } | |
7c673cae FG |
246 | } |
247 | } | |
248 | } | |
249 | return 0; | |
250 | } | |
251 | ||
28e407b8 AA |
252 | // ---------------------------------- |
253 | // OSDMap | |
7c673cae FG |
254 | |
255 | bool OSDMap::subtree_is_down(int id, set<int> *down_cache) const | |
256 | { | |
257 | if (id >= 0) | |
258 | return is_down(id); | |
259 | ||
260 | if (down_cache && | |
261 | down_cache->count(id)) { | |
262 | return true; | |
263 | } | |
264 | ||
265 | list<int> children; | |
266 | crush->get_children(id, &children); | |
267 | for (const auto &child : children) { | |
268 | if (!subtree_is_down(child, down_cache)) { | |
269 | return false; | |
270 | } | |
271 | } | |
272 | if (down_cache) { | |
273 | down_cache->insert(id); | |
274 | } | |
275 | return true; | |
276 | } | |
277 | ||
278 | bool OSDMap::containing_subtree_is_down(CephContext *cct, int id, int subtree_type, set<int> *down_cache) const | |
279 | { | |
280 | // use a stack-local down_cache if we didn't get one from the | |
281 | // caller. then at least this particular call will avoid duplicated | |
282 | // work. | |
283 | set<int> local_down_cache; | |
284 | if (!down_cache) { | |
285 | down_cache = &local_down_cache; | |
286 | } | |
287 | ||
288 | int current = id; | |
289 | while (true) { | |
290 | int type; | |
291 | if (current >= 0) { | |
292 | type = 0; | |
293 | } else { | |
294 | type = crush->get_bucket_type(current); | |
295 | } | |
11fdf7f2 | 296 | ceph_assert(type >= 0); |
7c673cae FG |
297 | |
298 | if (!subtree_is_down(current, down_cache)) { | |
299 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = false" << dendl; | |
300 | return false; | |
301 | } | |
302 | ||
303 | // is this a big enough subtree to be marked as down? | |
304 | if (type >= subtree_type) { | |
305 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = true ... " << type << " >= " << subtree_type << dendl; | |
306 | return true; | |
307 | } | |
308 | ||
309 | int r = crush->get_immediate_parent_id(current, ¤t); | |
310 | if (r < 0) { | |
311 | return false; | |
312 | } | |
313 | } | |
314 | } | |
315 | ||
224ce89b WB |
316 | bool OSDMap::subtree_type_is_down( |
317 | CephContext *cct, | |
318 | int id, | |
319 | int subtree_type, | |
320 | set<int> *down_in_osds, | |
321 | set<int> *up_in_osds, | |
322 | set<int> *subtree_up, | |
323 | unordered_map<int, set<int> > *subtree_type_down) const | |
31f18b77 FG |
324 | { |
325 | if (id >= 0) { | |
326 | bool is_down_ret = is_down(id); | |
327 | if (!is_out(id)) { | |
328 | if (is_down_ret) { | |
329 | down_in_osds->insert(id); | |
330 | } else { | |
331 | up_in_osds->insert(id); | |
332 | } | |
333 | } | |
334 | return is_down_ret; | |
335 | } | |
336 | ||
337 | if (subtree_type_down && | |
338 | (*subtree_type_down)[subtree_type].count(id)) { | |
339 | return true; | |
340 | } | |
341 | ||
342 | list<int> children; | |
343 | crush->get_children(id, &children); | |
344 | for (const auto &child : children) { | |
224ce89b WB |
345 | if (!subtree_type_is_down( |
346 | cct, child, crush->get_bucket_type(child), | |
347 | down_in_osds, up_in_osds, subtree_up, subtree_type_down)) { | |
31f18b77 FG |
348 | subtree_up->insert(id); |
349 | return false; | |
350 | } | |
351 | } | |
352 | if (subtree_type_down) { | |
353 | (*subtree_type_down)[subtree_type].insert(id); | |
354 | } | |
355 | return true; | |
356 | } | |
357 | ||
7c673cae FG |
358 | void OSDMap::Incremental::encode_client_old(bufferlist& bl) const |
359 | { | |
11fdf7f2 | 360 | using ceph::encode; |
7c673cae | 361 | __u16 v = 5; |
11fdf7f2 TL |
362 | encode(v, bl); |
363 | encode(fsid, bl); | |
364 | encode(epoch, bl); | |
365 | encode(modified, bl); | |
7c673cae | 366 | int32_t new_t = new_pool_max; |
11fdf7f2 TL |
367 | encode(new_t, bl); |
368 | encode(new_flags, bl); | |
369 | encode(fullmap, bl); | |
370 | encode(crush, bl); | |
7c673cae | 371 | |
11fdf7f2 TL |
372 | encode(new_max_osd, bl); |
373 | // for encode(new_pools, bl); | |
7c673cae | 374 | __u32 n = new_pools.size(); |
11fdf7f2 | 375 | encode(n, bl); |
7c673cae FG |
376 | for (const auto &new_pool : new_pools) { |
377 | n = new_pool.first; | |
11fdf7f2 TL |
378 | encode(n, bl); |
379 | encode(new_pool.second, bl, 0); | |
7c673cae | 380 | } |
11fdf7f2 | 381 | // for encode(new_pool_names, bl); |
7c673cae | 382 | n = new_pool_names.size(); |
11fdf7f2 | 383 | encode(n, bl); |
7c673cae FG |
384 | |
385 | for (const auto &new_pool_name : new_pool_names) { | |
386 | n = new_pool_name.first; | |
11fdf7f2 TL |
387 | encode(n, bl); |
388 | encode(new_pool_name.second, bl); | |
7c673cae | 389 | } |
11fdf7f2 | 390 | // for encode(old_pools, bl); |
7c673cae | 391 | n = old_pools.size(); |
11fdf7f2 | 392 | encode(n, bl); |
7c673cae FG |
393 | for (auto &old_pool : old_pools) { |
394 | n = old_pool; | |
11fdf7f2 | 395 | encode(n, bl); |
7c673cae | 396 | } |
11fdf7f2 | 397 | encode(new_up_client, bl, 0); |
31f18b77 FG |
398 | { |
399 | // legacy is map<int32_t,uint8_t> | |
400 | uint32_t n = new_state.size(); | |
11fdf7f2 | 401 | encode(n, bl); |
31f18b77 | 402 | for (auto p : new_state) { |
11fdf7f2 TL |
403 | encode(p.first, bl); |
404 | encode((uint8_t)p.second, bl); | |
31f18b77 FG |
405 | } |
406 | } | |
11fdf7f2 TL |
407 | encode(new_weight, bl); |
408 | // for encode(new_pg_temp, bl); | |
7c673cae | 409 | n = new_pg_temp.size(); |
11fdf7f2 | 410 | encode(n, bl); |
7c673cae FG |
411 | |
412 | for (const auto &pg_temp : new_pg_temp) { | |
413 | old_pg_t opg = pg_temp.first.get_old_pg(); | |
11fdf7f2 TL |
414 | encode(opg, bl); |
415 | encode(pg_temp.second, bl); | |
7c673cae FG |
416 | } |
417 | } | |
418 | ||
419 | void OSDMap::Incremental::encode_classic(bufferlist& bl, uint64_t features) const | |
420 | { | |
11fdf7f2 | 421 | using ceph::encode; |
7c673cae FG |
422 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
423 | encode_client_old(bl); | |
424 | return; | |
425 | } | |
426 | ||
427 | // base | |
428 | __u16 v = 6; | |
11fdf7f2 TL |
429 | encode(v, bl); |
430 | encode(fsid, bl); | |
431 | encode(epoch, bl); | |
432 | encode(modified, bl); | |
433 | encode(new_pool_max, bl); | |
434 | encode(new_flags, bl); | |
435 | encode(fullmap, bl); | |
436 | encode(crush, bl); | |
437 | ||
438 | encode(new_max_osd, bl); | |
439 | encode(new_pools, bl, features); | |
440 | encode(new_pool_names, bl); | |
441 | encode(old_pools, bl); | |
442 | encode(new_up_client, bl, features); | |
31f18b77 FG |
443 | { |
444 | uint32_t n = new_state.size(); | |
11fdf7f2 | 445 | encode(n, bl); |
31f18b77 | 446 | for (auto p : new_state) { |
11fdf7f2 TL |
447 | encode(p.first, bl); |
448 | encode((uint8_t)p.second, bl); | |
31f18b77 FG |
449 | } |
450 | } | |
11fdf7f2 TL |
451 | encode(new_weight, bl); |
452 | encode(new_pg_temp, bl); | |
7c673cae FG |
453 | |
454 | // extended | |
455 | __u16 ev = 10; | |
11fdf7f2 TL |
456 | encode(ev, bl); |
457 | encode(new_hb_back_up, bl, features); | |
458 | encode(new_up_thru, bl); | |
459 | encode(new_last_clean_interval, bl); | |
460 | encode(new_lost, bl); | |
461 | encode(new_blacklist, bl, features); | |
462 | encode(old_blacklist, bl, features); | |
463 | encode(new_up_cluster, bl, features); | |
464 | encode(cluster_snapshot, bl); | |
465 | encode(new_uuid, bl); | |
466 | encode(new_xinfo, bl); | |
467 | encode(new_hb_front_up, bl, features); | |
468 | } | |
469 | ||
470 | template<class T> | |
471 | static void encode_addrvec_map_as_addr(const T& m, bufferlist& bl, uint64_t f) | |
472 | { | |
473 | uint32_t n = m.size(); | |
474 | encode(n, bl); | |
475 | for (auto& i : m) { | |
476 | encode(i.first, bl); | |
477 | encode(i.second.legacy_addr(), bl, f); | |
478 | } | |
479 | } | |
480 | ||
481 | template<class T> | |
482 | static void encode_addrvec_pvec_as_addr(const T& m, bufferlist& bl, uint64_t f) | |
483 | { | |
484 | uint32_t n = m.size(); | |
485 | encode(n, bl); | |
486 | for (auto& i : m) { | |
487 | if (i) { | |
488 | encode(i->legacy_addr(), bl, f); | |
489 | } else { | |
490 | encode(entity_addr_t(), bl, f); | |
491 | } | |
492 | } | |
7c673cae FG |
493 | } |
494 | ||
11fdf7f2 TL |
495 | /* for a description of osdmap incremental versions, and when they were |
496 | * introduced, please refer to | |
497 | * doc/dev/osd_internals/osdmap_versions.txt | |
498 | */ | |
7c673cae FG |
499 | void OSDMap::Incremental::encode(bufferlist& bl, uint64_t features) const |
500 | { | |
11fdf7f2 | 501 | using ceph::encode; |
7c673cae FG |
502 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
503 | encode_classic(bl, features); | |
504 | return; | |
505 | } | |
506 | ||
507 | // only a select set of callers should *ever* be encoding new | |
508 | // OSDMaps. others should be passing around the canonical encoded | |
509 | // buffers from on high. select out those callers by passing in an | |
510 | // "impossible" feature bit. | |
11fdf7f2 | 511 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
512 | features &= ~CEPH_FEATURE_RESERVED; |
513 | ||
514 | size_t start_offset = bl.length(); | |
515 | size_t tail_offset; | |
11fdf7f2 TL |
516 | size_t crc_offset; |
517 | std::optional<buffer::list::contiguous_filler> crc_filler; | |
7c673cae FG |
518 | |
519 | // meta-encoding: how we include client-used and osd-specific data | |
520 | ENCODE_START(8, 7, bl); | |
521 | ||
522 | { | |
11fdf7f2 | 523 | uint8_t v = 8; |
7c673cae FG |
524 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
525 | v = 3; | |
11fdf7f2 TL |
526 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
527 | v = 5; | |
528 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
529 | v = 6; | |
7c673cae FG |
530 | } |
531 | ENCODE_START(v, 1, bl); // client-usable data | |
11fdf7f2 TL |
532 | encode(fsid, bl); |
533 | encode(epoch, bl); | |
534 | encode(modified, bl); | |
535 | encode(new_pool_max, bl); | |
536 | encode(new_flags, bl); | |
537 | encode(fullmap, bl); | |
538 | encode(crush, bl); | |
539 | ||
540 | encode(new_max_osd, bl); | |
541 | encode(new_pools, bl, features); | |
542 | encode(new_pool_names, bl); | |
543 | encode(old_pools, bl); | |
544 | if (v >= 7) { | |
545 | encode(new_up_client, bl, features); | |
546 | } else { | |
547 | encode_addrvec_map_as_addr(new_up_client, bl, features); | |
548 | } | |
31f18b77 | 549 | if (v >= 5) { |
11fdf7f2 | 550 | encode(new_state, bl); |
31f18b77 FG |
551 | } else { |
552 | uint32_t n = new_state.size(); | |
11fdf7f2 | 553 | encode(n, bl); |
31f18b77 | 554 | for (auto p : new_state) { |
11fdf7f2 TL |
555 | encode(p.first, bl); |
556 | encode((uint8_t)p.second, bl); | |
31f18b77 FG |
557 | } |
558 | } | |
11fdf7f2 TL |
559 | encode(new_weight, bl); |
560 | encode(new_pg_temp, bl); | |
561 | encode(new_primary_temp, bl); | |
562 | encode(new_primary_affinity, bl); | |
563 | encode(new_erasure_code_profiles, bl); | |
564 | encode(old_erasure_code_profiles, bl); | |
7c673cae | 565 | if (v >= 4) { |
11fdf7f2 TL |
566 | encode(new_pg_upmap, bl); |
567 | encode(old_pg_upmap, bl); | |
568 | encode(new_pg_upmap_items, bl); | |
569 | encode(old_pg_upmap_items, bl); | |
570 | } | |
571 | if (v >= 6) { | |
572 | encode(new_removed_snaps, bl); | |
573 | encode(new_purged_snaps, bl); | |
574 | } | |
575 | if (v >= 8) { | |
576 | encode(new_last_up_change, bl); | |
577 | encode(new_last_in_change, bl); | |
7c673cae FG |
578 | } |
579 | ENCODE_FINISH(bl); // client-usable data | |
580 | } | |
581 | ||
582 | { | |
81eedcae | 583 | uint8_t target_v = 9; |
7c673cae FG |
584 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
585 | target_v = 2; | |
11fdf7f2 TL |
586 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { |
587 | target_v = 6; | |
7c673cae FG |
588 | } |
589 | ENCODE_START(target_v, 1, bl); // extended, osd-only data | |
11fdf7f2 TL |
590 | if (target_v < 7) { |
591 | encode_addrvec_map_as_addr(new_hb_back_up, bl, features); | |
592 | } else { | |
593 | encode(new_hb_back_up, bl, features); | |
594 | } | |
595 | encode(new_up_thru, bl); | |
596 | encode(new_last_clean_interval, bl); | |
597 | encode(new_lost, bl); | |
598 | encode(new_blacklist, bl, features); | |
599 | encode(old_blacklist, bl, features); | |
600 | if (target_v < 7) { | |
601 | encode_addrvec_map_as_addr(new_up_cluster, bl, features); | |
602 | } else { | |
603 | encode(new_up_cluster, bl, features); | |
604 | } | |
605 | encode(cluster_snapshot, bl); | |
606 | encode(new_uuid, bl); | |
607 | encode(new_xinfo, bl); | |
608 | if (target_v < 7) { | |
609 | encode_addrvec_map_as_addr(new_hb_front_up, bl, features); | |
610 | } else { | |
611 | encode(new_hb_front_up, bl, features); | |
612 | } | |
613 | encode(features, bl); // NOTE: features arg, not the member | |
7c673cae | 614 | if (target_v >= 3) { |
11fdf7f2 TL |
615 | encode(new_nearfull_ratio, bl); |
616 | encode(new_full_ratio, bl); | |
617 | encode(new_backfillfull_ratio, bl); | |
31f18b77 FG |
618 | } |
619 | // 5 was string-based new_require_min_compat_client | |
620 | if (target_v >= 6) { | |
11fdf7f2 TL |
621 | encode(new_require_min_compat_client, bl); |
622 | encode(new_require_osd_release, bl); | |
7c673cae | 623 | } |
81eedcae TL |
624 | if (target_v >= 8) { |
625 | encode(new_crush_node_flags, bl); | |
626 | } | |
627 | if (target_v >= 9) { | |
628 | encode(new_device_class_flags, bl); | |
629 | } | |
7c673cae FG |
630 | ENCODE_FINISH(bl); // osd-only data |
631 | } | |
632 | ||
11fdf7f2 TL |
633 | crc_offset = bl.length(); |
634 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
635 | tail_offset = bl.length(); |
636 | ||
11fdf7f2 | 637 | encode(full_crc, bl); |
7c673cae FG |
638 | |
639 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
640 | ||
641 | // fill in crc | |
642 | bufferlist front; | |
11fdf7f2 | 643 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae FG |
644 | inc_crc = front.crc32c(-1); |
645 | bufferlist tail; | |
646 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); | |
647 | inc_crc = tail.crc32c(inc_crc); | |
648 | ceph_le32 crc_le; | |
649 | crc_le = inc_crc; | |
11fdf7f2 | 650 | crc_filler->copy_in(4u, (char*)&crc_le); |
7c673cae FG |
651 | have_crc = true; |
652 | } | |
653 | ||
11fdf7f2 | 654 | void OSDMap::Incremental::decode_classic(bufferlist::const_iterator &p) |
7c673cae | 655 | { |
11fdf7f2 | 656 | using ceph::decode; |
7c673cae FG |
657 | __u32 n, t; |
658 | // base | |
659 | __u16 v; | |
11fdf7f2 TL |
660 | decode(v, p); |
661 | decode(fsid, p); | |
662 | decode(epoch, p); | |
663 | decode(modified, p); | |
7c673cae | 664 | if (v == 4 || v == 5) { |
11fdf7f2 | 665 | decode(n, p); |
7c673cae FG |
666 | new_pool_max = n; |
667 | } else if (v >= 6) | |
11fdf7f2 TL |
668 | decode(new_pool_max, p); |
669 | decode(new_flags, p); | |
670 | decode(fullmap, p); | |
671 | decode(crush, p); | |
7c673cae | 672 | |
11fdf7f2 | 673 | decode(new_max_osd, p); |
7c673cae FG |
674 | if (v < 6) { |
675 | new_pools.clear(); | |
11fdf7f2 | 676 | decode(n, p); |
7c673cae | 677 | while (n--) { |
11fdf7f2 TL |
678 | decode(t, p); |
679 | decode(new_pools[t], p); | |
7c673cae FG |
680 | } |
681 | } else { | |
11fdf7f2 | 682 | decode(new_pools, p); |
7c673cae FG |
683 | } |
684 | if (v == 5) { | |
685 | new_pool_names.clear(); | |
11fdf7f2 | 686 | decode(n, p); |
7c673cae | 687 | while (n--) { |
11fdf7f2 TL |
688 | decode(t, p); |
689 | decode(new_pool_names[t], p); | |
7c673cae FG |
690 | } |
691 | } else if (v >= 6) { | |
11fdf7f2 | 692 | decode(new_pool_names, p); |
7c673cae FG |
693 | } |
694 | if (v < 6) { | |
695 | old_pools.clear(); | |
11fdf7f2 | 696 | decode(n, p); |
7c673cae | 697 | while (n--) { |
11fdf7f2 | 698 | decode(t, p); |
7c673cae FG |
699 | old_pools.insert(t); |
700 | } | |
701 | } else { | |
11fdf7f2 | 702 | decode(old_pools, p); |
7c673cae | 703 | } |
11fdf7f2 | 704 | decode(new_up_client, p); |
31f18b77 FG |
705 | { |
706 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 707 | decode(ns, p); |
31f18b77 FG |
708 | for (auto q : ns) { |
709 | new_state[q.first] = q.second; | |
710 | } | |
711 | } | |
11fdf7f2 | 712 | decode(new_weight, p); |
7c673cae FG |
713 | |
714 | if (v < 6) { | |
715 | new_pg_temp.clear(); | |
11fdf7f2 | 716 | decode(n, p); |
7c673cae FG |
717 | while (n--) { |
718 | old_pg_t opg; | |
719 | ::decode_raw(opg, p); | |
11fdf7f2 | 720 | decode(new_pg_temp[pg_t(opg)], p); |
7c673cae FG |
721 | } |
722 | } else { | |
11fdf7f2 | 723 | decode(new_pg_temp, p); |
7c673cae FG |
724 | } |
725 | ||
726 | // decode short map, too. | |
727 | if (v == 5 && p.end()) | |
728 | return; | |
729 | ||
730 | // extended | |
731 | __u16 ev = 0; | |
732 | if (v >= 5) | |
11fdf7f2 TL |
733 | decode(ev, p); |
734 | decode(new_hb_back_up, p); | |
7c673cae | 735 | if (v < 5) |
11fdf7f2 TL |
736 | decode(new_pool_names, p); |
737 | decode(new_up_thru, p); | |
738 | decode(new_last_clean_interval, p); | |
739 | decode(new_lost, p); | |
740 | decode(new_blacklist, p); | |
741 | decode(old_blacklist, p); | |
7c673cae | 742 | if (ev >= 6) |
11fdf7f2 | 743 | decode(new_up_cluster, p); |
7c673cae | 744 | if (ev >= 7) |
11fdf7f2 | 745 | decode(cluster_snapshot, p); |
7c673cae | 746 | if (ev >= 8) |
11fdf7f2 | 747 | decode(new_uuid, p); |
7c673cae | 748 | if (ev >= 9) |
11fdf7f2 | 749 | decode(new_xinfo, p); |
7c673cae | 750 | if (ev >= 10) |
11fdf7f2 | 751 | decode(new_hb_front_up, p); |
7c673cae FG |
752 | } |
753 | ||
11fdf7f2 TL |
754 | /* for a description of osdmap incremental versions, and when they were |
755 | * introduced, please refer to | |
756 | * doc/dev/osd_internals/osdmap_versions.txt | |
757 | */ | |
758 | void OSDMap::Incremental::decode(bufferlist::const_iterator& bl) | |
7c673cae | 759 | { |
11fdf7f2 | 760 | using ceph::decode; |
7c673cae FG |
761 | /** |
762 | * Older encodings of the Incremental had a single struct_v which | |
763 | * covered the whole encoding, and was prior to our modern | |
764 | * stuff which includes a compatv and a size. So if we see | |
765 | * a struct_v < 7, we must rewind to the beginning and use our | |
766 | * classic decoder. | |
767 | */ | |
768 | size_t start_offset = bl.get_off(); | |
769 | size_t tail_offset = 0; | |
770 | bufferlist crc_front, crc_tail; | |
771 | ||
772 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
773 | if (struct_v < 7) { | |
11fdf7f2 | 774 | bl.seek(start_offset); |
7c673cae FG |
775 | decode_classic(bl); |
776 | encode_features = 0; | |
777 | if (struct_v >= 6) | |
778 | encode_features = CEPH_FEATURE_PGID64; | |
779 | else | |
780 | encode_features = 0; | |
781 | return; | |
782 | } | |
783 | { | |
11fdf7f2 TL |
784 | DECODE_START(8, bl); // client-usable data |
785 | decode(fsid, bl); | |
786 | decode(epoch, bl); | |
787 | decode(modified, bl); | |
788 | decode(new_pool_max, bl); | |
789 | decode(new_flags, bl); | |
790 | decode(fullmap, bl); | |
791 | decode(crush, bl); | |
792 | ||
793 | decode(new_max_osd, bl); | |
794 | decode(new_pools, bl); | |
795 | decode(new_pool_names, bl); | |
796 | decode(old_pools, bl); | |
797 | decode(new_up_client, bl); | |
31f18b77 | 798 | if (struct_v >= 5) { |
11fdf7f2 | 799 | decode(new_state, bl); |
31f18b77 FG |
800 | } else { |
801 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 802 | decode(ns, bl); |
31f18b77 FG |
803 | for (auto q : ns) { |
804 | new_state[q.first] = q.second; | |
805 | } | |
806 | } | |
11fdf7f2 TL |
807 | decode(new_weight, bl); |
808 | decode(new_pg_temp, bl); | |
809 | decode(new_primary_temp, bl); | |
7c673cae | 810 | if (struct_v >= 2) |
11fdf7f2 | 811 | decode(new_primary_affinity, bl); |
7c673cae FG |
812 | else |
813 | new_primary_affinity.clear(); | |
814 | if (struct_v >= 3) { | |
11fdf7f2 TL |
815 | decode(new_erasure_code_profiles, bl); |
816 | decode(old_erasure_code_profiles, bl); | |
7c673cae FG |
817 | } else { |
818 | new_erasure_code_profiles.clear(); | |
819 | old_erasure_code_profiles.clear(); | |
820 | } | |
821 | if (struct_v >= 4) { | |
11fdf7f2 TL |
822 | decode(new_pg_upmap, bl); |
823 | decode(old_pg_upmap, bl); | |
824 | decode(new_pg_upmap_items, bl); | |
825 | decode(old_pg_upmap_items, bl); | |
826 | } | |
827 | if (struct_v >= 6) { | |
828 | decode(new_removed_snaps, bl); | |
829 | decode(new_purged_snaps, bl); | |
830 | } | |
831 | if (struct_v >= 8) { | |
832 | decode(new_last_up_change, bl); | |
833 | decode(new_last_in_change, bl); | |
7c673cae FG |
834 | } |
835 | DECODE_FINISH(bl); // client-usable data | |
836 | } | |
837 | ||
838 | { | |
81eedcae | 839 | DECODE_START(9, bl); // extended, osd-only data |
11fdf7f2 TL |
840 | decode(new_hb_back_up, bl); |
841 | decode(new_up_thru, bl); | |
842 | decode(new_last_clean_interval, bl); | |
843 | decode(new_lost, bl); | |
844 | decode(new_blacklist, bl); | |
845 | decode(old_blacklist, bl); | |
846 | decode(new_up_cluster, bl); | |
847 | decode(cluster_snapshot, bl); | |
848 | decode(new_uuid, bl); | |
849 | decode(new_xinfo, bl); | |
850 | decode(new_hb_front_up, bl); | |
7c673cae | 851 | if (struct_v >= 2) |
11fdf7f2 | 852 | decode(encode_features, bl); |
7c673cae FG |
853 | else |
854 | encode_features = CEPH_FEATURE_PGID64 | CEPH_FEATURE_OSDMAP_ENC; | |
855 | if (struct_v >= 3) { | |
11fdf7f2 TL |
856 | decode(new_nearfull_ratio, bl); |
857 | decode(new_full_ratio, bl); | |
7c673cae FG |
858 | } else { |
859 | new_nearfull_ratio = -1; | |
860 | new_full_ratio = -1; | |
861 | } | |
862 | if (struct_v >= 4) { | |
11fdf7f2 | 863 | decode(new_backfillfull_ratio, bl); |
7c673cae FG |
864 | } else { |
865 | new_backfillfull_ratio = -1; | |
866 | } | |
31f18b77 FG |
867 | if (struct_v == 5) { |
868 | string r; | |
11fdf7f2 | 869 | decode(r, bl); |
31f18b77 FG |
870 | if (r.length()) { |
871 | new_require_min_compat_client = ceph_release_from_name(r.c_str()); | |
872 | } | |
873 | } | |
874 | if (struct_v >= 6) { | |
11fdf7f2 TL |
875 | decode(new_require_min_compat_client, bl); |
876 | decode(new_require_osd_release, bl); | |
31f18b77 FG |
877 | } else { |
878 | if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_LUMINOUS)) { | |
879 | // only for compat with post-kraken pre-luminous test clusters | |
880 | new_require_osd_release = CEPH_RELEASE_LUMINOUS; | |
881 | new_flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
882 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_KRAKEN)) { | |
883 | new_require_osd_release = CEPH_RELEASE_KRAKEN; | |
884 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_JEWEL)) { | |
885 | new_require_osd_release = CEPH_RELEASE_JEWEL; | |
886 | } else { | |
887 | new_require_osd_release = -1; | |
888 | } | |
889 | } | |
81eedcae TL |
890 | if (struct_v >= 8) { |
891 | decode(new_crush_node_flags, bl); | |
892 | } | |
893 | if (struct_v >= 9) { | |
894 | decode(new_device_class_flags, bl); | |
895 | } | |
7c673cae FG |
896 | DECODE_FINISH(bl); // osd-only data |
897 | } | |
898 | ||
899 | if (struct_v >= 8) { | |
900 | have_crc = true; | |
901 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 902 | decode(inc_crc, bl); |
7c673cae | 903 | tail_offset = bl.get_off(); |
11fdf7f2 | 904 | decode(full_crc, bl); |
7c673cae FG |
905 | } else { |
906 | have_crc = false; | |
907 | full_crc = 0; | |
908 | inc_crc = 0; | |
909 | } | |
910 | ||
911 | DECODE_FINISH(bl); // wrapper | |
912 | ||
913 | if (have_crc) { | |
914 | // verify crc | |
915 | uint32_t actual = crc_front.crc32c(-1); | |
916 | if (tail_offset < bl.get_off()) { | |
917 | bufferlist tail; | |
918 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); | |
919 | actual = tail.crc32c(actual); | |
920 | } | |
921 | if (inc_crc != actual) { | |
922 | ostringstream ss; | |
923 | ss << "bad crc, actual " << actual << " != expected " << inc_crc; | |
924 | string s = ss.str(); | |
925 | throw buffer::malformed_input(s.c_str()); | |
926 | } | |
927 | } | |
928 | } | |
929 | ||
930 | void OSDMap::Incremental::dump(Formatter *f) const | |
931 | { | |
932 | f->dump_int("epoch", epoch); | |
933 | f->dump_stream("fsid") << fsid; | |
934 | f->dump_stream("modified") << modified; | |
11fdf7f2 TL |
935 | f->dump_stream("new_last_up_change") << new_last_up_change; |
936 | f->dump_stream("new_last_in_change") << new_last_in_change; | |
7c673cae FG |
937 | f->dump_int("new_pool_max", new_pool_max); |
938 | f->dump_int("new_flags", new_flags); | |
939 | f->dump_float("new_full_ratio", new_full_ratio); | |
940 | f->dump_float("new_nearfull_ratio", new_nearfull_ratio); | |
941 | f->dump_float("new_backfillfull_ratio", new_backfillfull_ratio); | |
31f18b77 FG |
942 | f->dump_int("new_require_min_compat_client", new_require_min_compat_client); |
943 | f->dump_int("new_require_osd_release", new_require_osd_release); | |
7c673cae FG |
944 | |
945 | if (fullmap.length()) { | |
946 | f->open_object_section("full_map"); | |
947 | OSDMap full; | |
948 | bufferlist fbl = fullmap; // kludge around constness. | |
11fdf7f2 | 949 | auto p = fbl.cbegin(); |
7c673cae FG |
950 | full.decode(p); |
951 | full.dump(f); | |
952 | f->close_section(); | |
953 | } | |
954 | if (crush.length()) { | |
955 | f->open_object_section("crush"); | |
956 | CrushWrapper c; | |
957 | bufferlist tbl = crush; // kludge around constness. | |
11fdf7f2 | 958 | auto p = tbl.cbegin(); |
7c673cae FG |
959 | c.decode(p); |
960 | c.dump(f); | |
961 | f->close_section(); | |
962 | } | |
963 | ||
964 | f->dump_int("new_max_osd", new_max_osd); | |
965 | ||
966 | f->open_array_section("new_pools"); | |
967 | ||
968 | for (const auto &new_pool : new_pools) { | |
969 | f->open_object_section("pool"); | |
970 | f->dump_int("pool", new_pool.first); | |
971 | new_pool.second.dump(f); | |
972 | f->close_section(); | |
973 | } | |
974 | f->close_section(); | |
975 | f->open_array_section("new_pool_names"); | |
976 | ||
977 | for (const auto &new_pool_name : new_pool_names) { | |
978 | f->open_object_section("pool_name"); | |
979 | f->dump_int("pool", new_pool_name.first); | |
980 | f->dump_string("name", new_pool_name.second); | |
981 | f->close_section(); | |
982 | } | |
983 | f->close_section(); | |
984 | f->open_array_section("old_pools"); | |
985 | ||
986 | for (const auto &old_pool : old_pools) | |
987 | f->dump_int("pool", old_pool); | |
988 | f->close_section(); | |
989 | ||
990 | f->open_array_section("new_up_osds"); | |
991 | ||
992 | for (const auto &upclient : new_up_client) { | |
993 | f->open_object_section("osd"); | |
994 | f->dump_int("osd", upclient.first); | |
11fdf7f2 TL |
995 | f->dump_stream("public_addr") << upclient.second.legacy_addr(); |
996 | f->dump_object("public_addrs", upclient.second); | |
997 | if (auto p = new_up_cluster.find(upclient.first); | |
998 | p != new_up_cluster.end()) { | |
999 | f->dump_stream("cluster_addr") << p->second.legacy_addr(); | |
1000 | f->dump_object("cluster_addrs", p->second); | |
1001 | } | |
1002 | if (auto p = new_hb_back_up.find(upclient.first); | |
1003 | p != new_hb_back_up.end()) { | |
1004 | f->dump_object("heartbeat_back_addrs", p->second); | |
1005 | } | |
1006 | if (auto p = new_hb_front_up.find(upclient.first); | |
1007 | p != new_hb_front_up.end()) { | |
1008 | f->dump_object("heartbeat_front_addrs", p->second); | |
1009 | } | |
7c673cae FG |
1010 | f->close_section(); |
1011 | } | |
1012 | f->close_section(); | |
1013 | ||
1014 | f->open_array_section("new_weight"); | |
1015 | ||
1016 | for (const auto &weight : new_weight) { | |
1017 | f->open_object_section("osd"); | |
1018 | f->dump_int("osd", weight.first); | |
1019 | f->dump_int("weight", weight.second); | |
1020 | f->close_section(); | |
1021 | } | |
1022 | f->close_section(); | |
1023 | ||
1024 | f->open_array_section("osd_state_xor"); | |
1025 | for (const auto &ns : new_state) { | |
1026 | f->open_object_section("osd"); | |
1027 | f->dump_int("osd", ns.first); | |
1028 | set<string> st; | |
1029 | calc_state_set(new_state.find(ns.first)->second, st); | |
1030 | f->open_array_section("state_xor"); | |
1031 | for (auto &state : st) | |
1032 | f->dump_string("state", state); | |
1033 | f->close_section(); | |
c07f9fc5 | 1034 | f->close_section(); |
7c673cae FG |
1035 | } |
1036 | f->close_section(); | |
1037 | ||
1038 | f->open_array_section("new_pg_temp"); | |
1039 | ||
1040 | for (const auto &pg_temp : new_pg_temp) { | |
1041 | f->open_object_section("pg"); | |
1042 | f->dump_stream("pgid") << pg_temp.first; | |
1043 | f->open_array_section("osds"); | |
1044 | ||
1045 | for (const auto &osd : pg_temp.second) | |
1046 | f->dump_int("osd", osd); | |
1047 | f->close_section(); | |
1048 | f->close_section(); | |
1049 | } | |
1050 | f->close_section(); | |
1051 | ||
1052 | f->open_array_section("primary_temp"); | |
1053 | ||
1054 | for (const auto &primary_temp : new_primary_temp) { | |
1055 | f->dump_stream("pgid") << primary_temp.first; | |
1056 | f->dump_int("osd", primary_temp.second); | |
1057 | } | |
1058 | f->close_section(); // primary_temp | |
1059 | ||
1060 | f->open_array_section("new_pg_upmap"); | |
1061 | for (auto& i : new_pg_upmap) { | |
1062 | f->open_object_section("mapping"); | |
1063 | f->dump_stream("pgid") << i.first; | |
1064 | f->open_array_section("osds"); | |
1065 | for (auto osd : i.second) { | |
1066 | f->dump_int("osd", osd); | |
1067 | } | |
1068 | f->close_section(); | |
1069 | f->close_section(); | |
1070 | } | |
1071 | f->close_section(); | |
1072 | f->open_array_section("old_pg_upmap"); | |
1073 | for (auto& i : old_pg_upmap) { | |
1074 | f->dump_stream("pgid") << i; | |
1075 | } | |
1076 | f->close_section(); | |
1077 | ||
1078 | f->open_array_section("new_pg_upmap_items"); | |
1079 | for (auto& i : new_pg_upmap_items) { | |
1080 | f->open_object_section("mapping"); | |
1081 | f->dump_stream("pgid") << i.first; | |
1082 | f->open_array_section("mappings"); | |
1083 | for (auto& p : i.second) { | |
1084 | f->open_object_section("mapping"); | |
1085 | f->dump_int("from", p.first); | |
1086 | f->dump_int("to", p.second); | |
1087 | f->close_section(); | |
1088 | } | |
1089 | f->close_section(); | |
1090 | f->close_section(); | |
1091 | } | |
1092 | f->close_section(); | |
1093 | f->open_array_section("old_pg_upmap_items"); | |
1094 | for (auto& i : old_pg_upmap_items) { | |
1095 | f->dump_stream("pgid") << i; | |
1096 | } | |
1097 | f->close_section(); | |
1098 | ||
1099 | f->open_array_section("new_up_thru"); | |
1100 | ||
1101 | for (const auto &up_thru : new_up_thru) { | |
1102 | f->open_object_section("osd"); | |
1103 | f->dump_int("osd", up_thru.first); | |
1104 | f->dump_int("up_thru", up_thru.second); | |
1105 | f->close_section(); | |
1106 | } | |
1107 | f->close_section(); | |
1108 | ||
1109 | f->open_array_section("new_lost"); | |
1110 | ||
1111 | for (const auto &lost : new_lost) { | |
1112 | f->open_object_section("osd"); | |
1113 | f->dump_int("osd", lost.first); | |
1114 | f->dump_int("epoch_lost", lost.second); | |
1115 | f->close_section(); | |
1116 | } | |
1117 | f->close_section(); | |
1118 | ||
1119 | f->open_array_section("new_last_clean_interval"); | |
1120 | ||
1121 | for (const auto &last_clean_interval : new_last_clean_interval) { | |
1122 | f->open_object_section("osd"); | |
1123 | f->dump_int("osd", last_clean_interval.first); | |
1124 | f->dump_int("first", last_clean_interval.second.first); | |
1125 | f->dump_int("last", last_clean_interval.second.second); | |
1126 | f->close_section(); | |
1127 | } | |
1128 | f->close_section(); | |
1129 | ||
1130 | f->open_array_section("new_blacklist"); | |
1131 | for (const auto &blist : new_blacklist) { | |
1132 | stringstream ss; | |
1133 | ss << blist.first; | |
1134 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1135 | } | |
1136 | f->close_section(); | |
1137 | f->open_array_section("old_blacklist"); | |
1138 | for (const auto &blist : old_blacklist) | |
1139 | f->dump_stream("addr") << blist; | |
1140 | f->close_section(); | |
1141 | ||
1142 | f->open_array_section("new_xinfo"); | |
1143 | for (const auto &xinfo : new_xinfo) { | |
1144 | f->open_object_section("xinfo"); | |
1145 | f->dump_int("osd", xinfo.first); | |
1146 | xinfo.second.dump(f); | |
1147 | f->close_section(); | |
1148 | } | |
1149 | f->close_section(); | |
1150 | ||
1151 | if (cluster_snapshot.size()) | |
1152 | f->dump_string("cluster_snapshot", cluster_snapshot); | |
1153 | ||
1154 | f->open_array_section("new_uuid"); | |
1155 | for (const auto &uuid : new_uuid) { | |
1156 | f->open_object_section("osd"); | |
1157 | f->dump_int("osd", uuid.first); | |
1158 | f->dump_stream("uuid") << uuid.second; | |
1159 | f->close_section(); | |
1160 | } | |
1161 | f->close_section(); | |
1162 | ||
1163 | OSDMap::dump_erasure_code_profiles(new_erasure_code_profiles, f); | |
1164 | f->open_array_section("old_erasure_code_profiles"); | |
1165 | for (const auto &erasure_code_profile : old_erasure_code_profiles) { | |
1166 | f->dump_string("old", erasure_code_profile.c_str()); | |
1167 | } | |
1168 | f->close_section(); | |
11fdf7f2 TL |
1169 | |
1170 | f->open_array_section("new_removed_snaps"); | |
1171 | for (auto& p : new_removed_snaps) { | |
1172 | f->open_object_section("pool"); | |
1173 | f->dump_int("pool", p.first); | |
1174 | f->open_array_section("snaps"); | |
1175 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1176 | f->open_object_section("interval"); | |
1177 | f->dump_unsigned("begin", q.get_start()); | |
1178 | f->dump_unsigned("length", q.get_len()); | |
1179 | f->close_section(); | |
1180 | } | |
1181 | f->close_section(); | |
1182 | f->close_section(); | |
1183 | } | |
1184 | f->close_section(); | |
1185 | f->open_array_section("new_purged_snaps"); | |
1186 | for (auto& p : new_purged_snaps) { | |
1187 | f->open_object_section("pool"); | |
1188 | f->dump_int("pool", p.first); | |
1189 | f->open_array_section("snaps"); | |
1190 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1191 | f->open_object_section("interval"); | |
1192 | f->dump_unsigned("begin", q.get_start()); | |
1193 | f->dump_unsigned("length", q.get_len()); | |
1194 | f->close_section(); | |
1195 | } | |
1196 | f->close_section(); | |
1197 | f->close_section(); | |
1198 | } | |
81eedcae TL |
1199 | f->open_array_section("new_crush_node_flags"); |
1200 | for (auto& i : new_crush_node_flags) { | |
1201 | f->open_object_section("node"); | |
1202 | f->dump_int("id", i.first); | |
1203 | set<string> st; | |
1204 | calc_state_set(i.second, st); | |
1205 | for (auto& j : st) { | |
1206 | f->dump_string("flag", j); | |
1207 | } | |
1208 | f->close_section(); | |
1209 | } | |
1210 | f->close_section(); | |
1211 | f->open_array_section("new_device_class_flags"); | |
1212 | for (auto& i : new_device_class_flags) { | |
1213 | f->open_object_section("device_class"); | |
1214 | f->dump_int("id", i.first); | |
1215 | set<string> st; | |
1216 | calc_state_set(i.second, st); | |
1217 | for (auto& j : st) { | |
1218 | f->dump_string("flag", j); | |
1219 | } | |
1220 | f->close_section(); | |
1221 | } | |
1222 | f->close_section(); | |
11fdf7f2 | 1223 | f->close_section(); |
7c673cae FG |
1224 | } |
1225 | ||
1226 | void OSDMap::Incremental::generate_test_instances(list<Incremental*>& o) | |
1227 | { | |
1228 | o.push_back(new Incremental); | |
1229 | } | |
1230 | ||
1231 | // ---------------------------------- | |
1232 | // OSDMap | |
1233 | ||
1234 | void OSDMap::set_epoch(epoch_t e) | |
1235 | { | |
1236 | epoch = e; | |
1237 | for (auto &pool : pools) | |
1238 | pool.second.last_change = e; | |
1239 | } | |
1240 | ||
11fdf7f2 | 1241 | bool OSDMap::is_blacklisted(const entity_addr_t& orig) const |
7c673cae | 1242 | { |
11fdf7f2 | 1243 | if (blacklist.empty()) { |
7c673cae | 1244 | return false; |
11fdf7f2 TL |
1245 | } |
1246 | ||
1247 | // all blacklist entries are type ANY for nautilus+ | |
1248 | // FIXME: avoid this copy! | |
1249 | entity_addr_t a = orig; | |
1250 | if (require_osd_release < CEPH_RELEASE_NAUTILUS) { | |
1251 | a.set_type(entity_addr_t::TYPE_LEGACY); | |
1252 | } else { | |
1253 | a.set_type(entity_addr_t::TYPE_ANY); | |
1254 | } | |
7c673cae FG |
1255 | |
1256 | // this specific instance? | |
11fdf7f2 | 1257 | if (blacklist.count(a)) { |
7c673cae | 1258 | return true; |
11fdf7f2 | 1259 | } |
7c673cae FG |
1260 | |
1261 | // is entire ip blacklisted? | |
1262 | if (a.is_ip()) { | |
11fdf7f2 TL |
1263 | a.set_port(0); |
1264 | a.set_nonce(0); | |
1265 | if (blacklist.count(a)) { | |
1266 | return true; | |
1267 | } | |
1268 | } | |
1269 | ||
1270 | return false; | |
1271 | } | |
1272 | ||
1273 | bool OSDMap::is_blacklisted(const entity_addrvec_t& av) const | |
1274 | { | |
1275 | if (blacklist.empty()) | |
1276 | return false; | |
1277 | ||
1278 | for (auto& a : av.v) { | |
1279 | if (is_blacklisted(a)) { | |
7c673cae FG |
1280 | return true; |
1281 | } | |
1282 | } | |
1283 | ||
1284 | return false; | |
1285 | } | |
1286 | ||
1287 | void OSDMap::get_blacklist(list<pair<entity_addr_t,utime_t> > *bl) const | |
1288 | { | |
1289 | std::copy(blacklist.begin(), blacklist.end(), std::back_inserter(*bl)); | |
1290 | } | |
1291 | ||
31f18b77 FG |
1292 | void OSDMap::get_blacklist(std::set<entity_addr_t> *bl) const |
1293 | { | |
1294 | for (const auto &i : blacklist) { | |
1295 | bl->insert(i.first); | |
1296 | } | |
1297 | } | |
1298 | ||
7c673cae FG |
1299 | void OSDMap::set_max_osd(int m) |
1300 | { | |
1301 | int o = max_osd; | |
1302 | max_osd = m; | |
1303 | osd_state.resize(m); | |
1304 | osd_weight.resize(m); | |
1305 | for (; o<max_osd; o++) { | |
1306 | osd_state[o] = 0; | |
1307 | osd_weight[o] = CEPH_OSD_OUT; | |
1308 | } | |
1309 | osd_info.resize(m); | |
1310 | osd_xinfo.resize(m); | |
11fdf7f2 TL |
1311 | osd_addrs->client_addrs.resize(m); |
1312 | osd_addrs->cluster_addrs.resize(m); | |
1313 | osd_addrs->hb_back_addrs.resize(m); | |
1314 | osd_addrs->hb_front_addrs.resize(m); | |
7c673cae FG |
1315 | osd_uuid->resize(m); |
1316 | if (osd_primary_affinity) | |
1317 | osd_primary_affinity->resize(m, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
1318 | ||
1319 | calc_num_osds(); | |
1320 | } | |
1321 | ||
1322 | int OSDMap::calc_num_osds() | |
1323 | { | |
1324 | num_osd = 0; | |
1325 | num_up_osd = 0; | |
1326 | num_in_osd = 0; | |
1327 | for (int i=0; i<max_osd; i++) { | |
1328 | if (osd_state[i] & CEPH_OSD_EXISTS) { | |
1329 | ++num_osd; | |
1330 | if (osd_state[i] & CEPH_OSD_UP) { | |
1331 | ++num_up_osd; | |
1332 | } | |
1333 | if (get_weight(i) != CEPH_OSD_OUT) { | |
1334 | ++num_in_osd; | |
1335 | } | |
1336 | } | |
1337 | } | |
1338 | return num_osd; | |
1339 | } | |
1340 | ||
3efd9988 FG |
1341 | void OSDMap::get_full_pools(CephContext *cct, |
1342 | set<int64_t> *full, | |
1343 | set<int64_t> *backfillfull, | |
1344 | set<int64_t> *nearfull) const | |
7c673cae | 1345 | { |
11fdf7f2 TL |
1346 | ceph_assert(full); |
1347 | ceph_assert(backfillfull); | |
1348 | ceph_assert(nearfull); | |
3efd9988 FG |
1349 | full->clear(); |
1350 | backfillfull->clear(); | |
1351 | nearfull->clear(); | |
1352 | ||
1353 | vector<int> full_osds; | |
1354 | vector<int> backfillfull_osds; | |
1355 | vector<int> nearfull_osds; | |
7c673cae FG |
1356 | for (int i = 0; i < max_osd; ++i) { |
1357 | if (exists(i) && is_up(i) && is_in(i)) { | |
1358 | if (osd_state[i] & CEPH_OSD_FULL) | |
3efd9988 | 1359 | full_osds.push_back(i); |
7c673cae | 1360 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) |
3efd9988 | 1361 | backfillfull_osds.push_back(i); |
7c673cae | 1362 | else if (osd_state[i] & CEPH_OSD_NEARFULL) |
3efd9988 | 1363 | nearfull_osds.push_back(i); |
7c673cae FG |
1364 | } |
1365 | } | |
3efd9988 FG |
1366 | |
1367 | for (auto i: full_osds) { | |
1368 | get_pool_ids_by_osd(cct, i, full); | |
1369 | } | |
1370 | for (auto i: backfillfull_osds) { | |
1371 | get_pool_ids_by_osd(cct, i, backfillfull); | |
1372 | } | |
1373 | for (auto i: nearfull_osds) { | |
1374 | get_pool_ids_by_osd(cct, i, nearfull); | |
1375 | } | |
7c673cae FG |
1376 | } |
1377 | ||
31f18b77 FG |
1378 | void OSDMap::get_full_osd_counts(set<int> *full, set<int> *backfill, |
1379 | set<int> *nearfull) const | |
1380 | { | |
1381 | full->clear(); | |
1382 | backfill->clear(); | |
1383 | nearfull->clear(); | |
1384 | for (int i = 0; i < max_osd; ++i) { | |
1385 | if (exists(i) && is_up(i) && is_in(i)) { | |
1386 | if (osd_state[i] & CEPH_OSD_FULL) | |
1387 | full->emplace(i); | |
1388 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1389 | backfill->emplace(i); | |
1390 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1391 | nearfull->emplace(i); | |
1392 | } | |
1393 | } | |
1394 | } | |
1395 | ||
7c673cae FG |
1396 | void OSDMap::get_all_osds(set<int32_t>& ls) const |
1397 | { | |
1398 | for (int i=0; i<max_osd; i++) | |
1399 | if (exists(i)) | |
1400 | ls.insert(i); | |
1401 | } | |
1402 | ||
1403 | void OSDMap::get_up_osds(set<int32_t>& ls) const | |
1404 | { | |
1405 | for (int i = 0; i < max_osd; i++) { | |
1406 | if (is_up(i)) | |
1407 | ls.insert(i); | |
1408 | } | |
1409 | } | |
1410 | ||
81eedcae | 1411 | void OSDMap::get_out_existing_osds(set<int32_t>& ls) const |
31f18b77 FG |
1412 | { |
1413 | for (int i = 0; i < max_osd; i++) { | |
81eedcae | 1414 | if (exists(i) && get_weight(i) == CEPH_OSD_OUT) |
31f18b77 FG |
1415 | ls.insert(i); |
1416 | } | |
1417 | } | |
1418 | ||
11fdf7f2 TL |
1419 | void OSDMap::get_flag_set(set<string> *flagset) const |
1420 | { | |
1421 | for (unsigned i = 0; i < sizeof(flags) * 8; ++i) { | |
1422 | if (flags & (1<<i)) { | |
1423 | flagset->insert(get_flag_string(flags & (1<<i))); | |
1424 | } | |
1425 | } | |
1426 | } | |
1427 | ||
7c673cae FG |
1428 | void OSDMap::calc_state_set(int state, set<string>& st) |
1429 | { | |
1430 | unsigned t = state; | |
1431 | for (unsigned s = 1; t; s <<= 1) { | |
1432 | if (t & s) { | |
1433 | t &= ~s; | |
1434 | st.insert(ceph_osd_state_name(s)); | |
1435 | } | |
1436 | } | |
1437 | } | |
1438 | ||
1439 | void OSDMap::adjust_osd_weights(const map<int,double>& weights, Incremental& inc) const | |
1440 | { | |
1441 | float max = 0; | |
1442 | for (const auto &weight : weights) { | |
1443 | if (weight.second > max) | |
1444 | max = weight.second; | |
1445 | } | |
1446 | ||
1447 | for (const auto &weight : weights) { | |
1448 | inc.new_weight[weight.first] = (unsigned)((weight.second / max) * CEPH_OSD_IN); | |
1449 | } | |
1450 | } | |
1451 | ||
1452 | int OSDMap::identify_osd(const entity_addr_t& addr) const | |
1453 | { | |
1454 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1455 | if (exists(i) && (get_addrs(i).contains(addr) || |
1456 | get_cluster_addrs(i).contains(addr))) | |
7c673cae FG |
1457 | return i; |
1458 | return -1; | |
1459 | } | |
1460 | ||
1461 | int OSDMap::identify_osd(const uuid_d& u) const | |
1462 | { | |
1463 | for (int i=0; i<max_osd; i++) | |
1464 | if (exists(i) && get_uuid(i) == u) | |
1465 | return i; | |
1466 | return -1; | |
1467 | } | |
1468 | ||
1469 | int OSDMap::identify_osd_on_all_channels(const entity_addr_t& addr) const | |
1470 | { | |
1471 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1472 | if (exists(i) && (get_addrs(i).contains(addr) || |
1473 | get_cluster_addrs(i).contains(addr) || | |
1474 | get_hb_back_addrs(i).contains(addr) || | |
1475 | get_hb_front_addrs(i).contains(addr))) | |
7c673cae FG |
1476 | return i; |
1477 | return -1; | |
1478 | } | |
1479 | ||
1480 | int OSDMap::find_osd_on_ip(const entity_addr_t& ip) const | |
1481 | { | |
1482 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1483 | if (exists(i) && (get_addrs(i).is_same_host(ip) || |
1484 | get_cluster_addrs(i).is_same_host(ip))) | |
7c673cae FG |
1485 | return i; |
1486 | return -1; | |
1487 | } | |
1488 | ||
1489 | ||
1490 | uint64_t OSDMap::get_features(int entity_type, uint64_t *pmask) const | |
1491 | { | |
1492 | uint64_t features = 0; // things we actually have | |
1493 | uint64_t mask = 0; // things we could have | |
1494 | ||
1495 | if (crush->has_nondefault_tunables()) | |
1496 | features |= CEPH_FEATURE_CRUSH_TUNABLES; | |
1497 | if (crush->has_nondefault_tunables2()) | |
1498 | features |= CEPH_FEATURE_CRUSH_TUNABLES2; | |
1499 | if (crush->has_nondefault_tunables3()) | |
1500 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1501 | if (crush->has_v4_buckets()) | |
1502 | features |= CEPH_FEATURE_CRUSH_V4; | |
1503 | if (crush->has_nondefault_tunables5()) | |
1504 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
c07f9fc5 FG |
1505 | if (crush->has_incompat_choose_args()) { |
1506 | features |= CEPH_FEATUREMASK_CRUSH_CHOOSE_ARGS; | |
1507 | } | |
7c673cae FG |
1508 | mask |= CEPH_FEATURES_CRUSH; |
1509 | ||
1510 | if (!pg_upmap.empty() || !pg_upmap_items.empty()) | |
1511 | features |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1512 | mask |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1513 | ||
1514 | for (auto &pool: pools) { | |
1515 | if (pool.second.has_flag(pg_pool_t::FLAG_HASHPSPOOL)) { | |
1516 | features |= CEPH_FEATURE_OSDHASHPSPOOL; | |
1517 | } | |
7c673cae FG |
1518 | if (!pool.second.tiers.empty() || |
1519 | pool.second.is_tier()) { | |
1520 | features |= CEPH_FEATURE_OSD_CACHEPOOL; | |
1521 | } | |
31f18b77 | 1522 | int ruleid = crush->find_rule(pool.second.get_crush_rule(), |
7c673cae FG |
1523 | pool.second.get_type(), |
1524 | pool.second.get_size()); | |
1525 | if (ruleid >= 0) { | |
1526 | if (crush->is_v2_rule(ruleid)) | |
1527 | features |= CEPH_FEATURE_CRUSH_V2; | |
1528 | if (crush->is_v3_rule(ruleid)) | |
1529 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1530 | if (crush->is_v5_rule(ruleid)) | |
1531 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
1532 | } | |
1533 | } | |
7c673cae | 1534 | mask |= CEPH_FEATURE_OSDHASHPSPOOL | CEPH_FEATURE_OSD_CACHEPOOL; |
7c673cae FG |
1535 | |
1536 | if (osd_primary_affinity) { | |
1537 | for (int i = 0; i < max_osd; ++i) { | |
1538 | if ((*osd_primary_affinity)[i] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
1539 | features |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1540 | break; | |
1541 | } | |
1542 | } | |
1543 | } | |
1544 | mask |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1545 | ||
1546 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1547 | const uint64_t jewel_features = CEPH_FEATURE_SERVER_JEWEL; | |
31f18b77 | 1548 | if (require_osd_release >= CEPH_RELEASE_JEWEL) { |
7c673cae FG |
1549 | features |= jewel_features; |
1550 | } | |
1551 | mask |= jewel_features; | |
1552 | ||
1553 | const uint64_t kraken_features = CEPH_FEATUREMASK_SERVER_KRAKEN | |
1554 | | CEPH_FEATURE_MSG_ADDR2; | |
31f18b77 | 1555 | if (require_osd_release >= CEPH_RELEASE_KRAKEN) { |
7c673cae FG |
1556 | features |= kraken_features; |
1557 | } | |
1558 | mask |= kraken_features; | |
1559 | } | |
1560 | ||
11fdf7f2 TL |
1561 | if (require_min_compat_client >= CEPH_RELEASE_NAUTILUS) { |
1562 | // if min_compat_client is >= nautilus, require v2 cephx signatures | |
1563 | // from everyone | |
1564 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
1565 | } else if (require_osd_release >= CEPH_RELEASE_NAUTILUS && | |
1566 | entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1567 | // if osds are >= nautilus, at least require the signatures from them | |
1568 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
1569 | } | |
1570 | mask |= CEPH_FEATUREMASK_CEPHX_V2; | |
1571 | ||
7c673cae FG |
1572 | if (pmask) |
1573 | *pmask = mask; | |
1574 | return features; | |
1575 | } | |
1576 | ||
31f18b77 | 1577 | uint8_t OSDMap::get_min_compat_client() const |
7c673cae FG |
1578 | { |
1579 | uint64_t f = get_features(CEPH_ENTITY_TYPE_CLIENT, nullptr); | |
1580 | ||
1581 | if (HAVE_FEATURE(f, OSDMAP_PG_UPMAP) || // v12.0.0-1733-g27d6f43 | |
31f18b77 FG |
1582 | HAVE_FEATURE(f, CRUSH_CHOOSE_ARGS)) { // v12.0.1-2172-gef1ef28 |
1583 | return CEPH_RELEASE_LUMINOUS; // v12.2.0 | |
7c673cae FG |
1584 | } |
1585 | if (HAVE_FEATURE(f, CRUSH_TUNABLES5)) { // v10.0.0-612-g043a737 | |
31f18b77 | 1586 | return CEPH_RELEASE_JEWEL; // v10.2.0 |
7c673cae FG |
1587 | } |
1588 | if (HAVE_FEATURE(f, CRUSH_V4)) { // v0.91-678-g325fc56 | |
31f18b77 | 1589 | return CEPH_RELEASE_HAMMER; // v0.94.0 |
7c673cae FG |
1590 | } |
1591 | if (HAVE_FEATURE(f, OSD_PRIMARY_AFFINITY) || // v0.76-553-gf825624 | |
1592 | HAVE_FEATURE(f, CRUSH_TUNABLES3) || // v0.76-395-ge20a55d | |
7c673cae | 1593 | HAVE_FEATURE(f, OSD_CACHEPOOL)) { // v0.67-401-gb91c1c5 |
31f18b77 | 1594 | return CEPH_RELEASE_FIREFLY; // v0.80.0 |
7c673cae FG |
1595 | } |
1596 | if (HAVE_FEATURE(f, CRUSH_TUNABLES2) || // v0.54-684-g0cc47ff | |
1597 | HAVE_FEATURE(f, OSDHASHPSPOOL)) { // v0.57-398-g8cc2b0f | |
31f18b77 | 1598 | return CEPH_RELEASE_DUMPLING; // v0.67.0 |
7c673cae FG |
1599 | } |
1600 | if (HAVE_FEATURE(f, CRUSH_TUNABLES)) { // v0.48argonaut-206-g6f381af | |
31f18b77 | 1601 | return CEPH_RELEASE_ARGONAUT; // v0.48argonaut-206-g6f381af |
7c673cae | 1602 | } |
31f18b77 | 1603 | return CEPH_RELEASE_ARGONAUT; // v0.48argonaut-206-g6f381af |
7c673cae FG |
1604 | } |
1605 | ||
11fdf7f2 TL |
1606 | uint8_t OSDMap::get_require_min_compat_client() const |
1607 | { | |
1608 | return require_min_compat_client; | |
1609 | } | |
1610 | ||
7c673cae FG |
1611 | void OSDMap::_calc_up_osd_features() |
1612 | { | |
1613 | bool first = true; | |
1614 | cached_up_osd_features = 0; | |
1615 | for (int osd = 0; osd < max_osd; ++osd) { | |
1616 | if (!is_up(osd)) | |
1617 | continue; | |
1618 | const osd_xinfo_t &xi = get_xinfo(osd); | |
3efd9988 FG |
1619 | if (xi.features == 0) |
1620 | continue; // bogus xinfo, maybe #20751 or similar, skipping | |
7c673cae FG |
1621 | if (first) { |
1622 | cached_up_osd_features = xi.features; | |
1623 | first = false; | |
1624 | } else { | |
1625 | cached_up_osd_features &= xi.features; | |
1626 | } | |
1627 | } | |
1628 | } | |
1629 | ||
1630 | uint64_t OSDMap::get_up_osd_features() const | |
1631 | { | |
1632 | return cached_up_osd_features; | |
1633 | } | |
1634 | ||
1635 | void OSDMap::dedup(const OSDMap *o, OSDMap *n) | |
1636 | { | |
11fdf7f2 | 1637 | using ceph::encode; |
7c673cae FG |
1638 | if (o->epoch == n->epoch) |
1639 | return; | |
1640 | ||
1641 | int diff = 0; | |
1642 | ||
1643 | // do addrs match? | |
1644 | if (o->max_osd != n->max_osd) | |
1645 | diff++; | |
1646 | for (int i = 0; i < o->max_osd && i < n->max_osd; i++) { | |
11fdf7f2 TL |
1647 | if ( n->osd_addrs->client_addrs[i] && o->osd_addrs->client_addrs[i] && |
1648 | *n->osd_addrs->client_addrs[i] == *o->osd_addrs->client_addrs[i]) | |
1649 | n->osd_addrs->client_addrs[i] = o->osd_addrs->client_addrs[i]; | |
7c673cae FG |
1650 | else |
1651 | diff++; | |
11fdf7f2 TL |
1652 | if ( n->osd_addrs->cluster_addrs[i] && o->osd_addrs->cluster_addrs[i] && |
1653 | *n->osd_addrs->cluster_addrs[i] == *o->osd_addrs->cluster_addrs[i]) | |
1654 | n->osd_addrs->cluster_addrs[i] = o->osd_addrs->cluster_addrs[i]; | |
7c673cae FG |
1655 | else |
1656 | diff++; | |
11fdf7f2 TL |
1657 | if ( n->osd_addrs->hb_back_addrs[i] && o->osd_addrs->hb_back_addrs[i] && |
1658 | *n->osd_addrs->hb_back_addrs[i] == *o->osd_addrs->hb_back_addrs[i]) | |
1659 | n->osd_addrs->hb_back_addrs[i] = o->osd_addrs->hb_back_addrs[i]; | |
7c673cae FG |
1660 | else |
1661 | diff++; | |
11fdf7f2 TL |
1662 | if ( n->osd_addrs->hb_front_addrs[i] && o->osd_addrs->hb_front_addrs[i] && |
1663 | *n->osd_addrs->hb_front_addrs[i] == *o->osd_addrs->hb_front_addrs[i]) | |
1664 | n->osd_addrs->hb_front_addrs[i] = o->osd_addrs->hb_front_addrs[i]; | |
7c673cae FG |
1665 | else |
1666 | diff++; | |
1667 | } | |
1668 | if (diff == 0) { | |
1669 | // zoinks, no differences at all! | |
1670 | n->osd_addrs = o->osd_addrs; | |
1671 | } | |
1672 | ||
1673 | // does crush match? | |
1674 | bufferlist oc, nc; | |
11fdf7f2 TL |
1675 | encode(*o->crush, oc, CEPH_FEATURES_SUPPORTED_DEFAULT); |
1676 | encode(*n->crush, nc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
7c673cae FG |
1677 | if (oc.contents_equal(nc)) { |
1678 | n->crush = o->crush; | |
1679 | } | |
1680 | ||
1681 | // does pg_temp match? | |
31f18b77 FG |
1682 | if (*o->pg_temp == *n->pg_temp) |
1683 | n->pg_temp = o->pg_temp; | |
7c673cae FG |
1684 | |
1685 | // does primary_temp match? | |
1686 | if (o->primary_temp->size() == n->primary_temp->size()) { | |
1687 | if (*o->primary_temp == *n->primary_temp) | |
1688 | n->primary_temp = o->primary_temp; | |
1689 | } | |
1690 | ||
1691 | // do uuids match? | |
1692 | if (o->osd_uuid->size() == n->osd_uuid->size() && | |
1693 | *o->osd_uuid == *n->osd_uuid) | |
1694 | n->osd_uuid = o->osd_uuid; | |
1695 | } | |
1696 | ||
1697 | void OSDMap::clean_temps(CephContext *cct, | |
11fdf7f2 TL |
1698 | const OSDMap& oldmap, |
1699 | const OSDMap& nextmap, | |
1700 | Incremental *pending_inc) | |
7c673cae FG |
1701 | { |
1702 | ldout(cct, 10) << __func__ << dendl; | |
7c673cae | 1703 | |
11fdf7f2 | 1704 | for (auto pg : *nextmap.pg_temp) { |
7c673cae FG |
1705 | // if pool does not exist, remove any existing pg_temps associated with |
1706 | // it. we don't care about pg_temps on the pending_inc either; if there | |
1707 | // are new_pg_temp entries on the pending, clear them out just as well. | |
11fdf7f2 | 1708 | if (!nextmap.have_pg_pool(pg.first.pool())) { |
7c673cae FG |
1709 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first |
1710 | << " for nonexistent pool " << pg.first.pool() << dendl; | |
1711 | pending_inc->new_pg_temp[pg.first].clear(); | |
1712 | continue; | |
1713 | } | |
1714 | // all osds down? | |
1715 | unsigned num_up = 0; | |
1716 | for (auto o : pg.second) { | |
11fdf7f2 | 1717 | if (!nextmap.is_down(o)) { |
7c673cae FG |
1718 | ++num_up; |
1719 | break; | |
1720 | } | |
1721 | } | |
1722 | if (num_up == 0) { | |
1723 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1724 | << " with all down osds" << pg.second << dendl; | |
1725 | pending_inc->new_pg_temp[pg.first].clear(); | |
1726 | continue; | |
1727 | } | |
1728 | // redundant pg_temp? | |
1729 | vector<int> raw_up; | |
1730 | int primary; | |
11fdf7f2 | 1731 | nextmap.pg_to_raw_up(pg.first, &raw_up, &primary); |
91327a77 | 1732 | bool remove = false; |
11fdf7f2 | 1733 | if (raw_up == pg.second) { |
7c673cae FG |
1734 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1735 | << pg.second << " that matches raw_up mapping" << dendl; | |
91327a77 AA |
1736 | remove = true; |
1737 | } | |
1738 | // oversized pg_temp? | |
11fdf7f2 | 1739 | if (pg.second.size() > nextmap.get_pg_pool(pg.first.pool())->get_size()) { |
91327a77 AA |
1740 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1741 | << pg.second << " exceeds pool size" << dendl; | |
1742 | remove = true; | |
1743 | } | |
1744 | if (remove) { | |
11fdf7f2 | 1745 | if (oldmap.pg_temp->count(pg.first)) |
7c673cae FG |
1746 | pending_inc->new_pg_temp[pg.first].clear(); |
1747 | else | |
1748 | pending_inc->new_pg_temp.erase(pg.first); | |
1749 | } | |
1750 | } | |
1751 | ||
11fdf7f2 | 1752 | for (auto &pg : *nextmap.primary_temp) { |
7c673cae | 1753 | // primary down? |
11fdf7f2 | 1754 | if (nextmap.is_down(pg.second)) { |
7c673cae FG |
1755 | ldout(cct, 10) << __func__ << " removing primary_temp " << pg.first |
1756 | << " to down " << pg.second << dendl; | |
1757 | pending_inc->new_primary_temp[pg.first] = -1; | |
1758 | continue; | |
1759 | } | |
1760 | // redundant primary_temp? | |
1761 | vector<int> real_up, templess_up; | |
1762 | int real_primary, templess_primary; | |
1763 | pg_t pgid = pg.first; | |
11fdf7f2 TL |
1764 | nextmap.pg_to_acting_osds(pgid, &real_up, &real_primary); |
1765 | nextmap.pg_to_raw_up(pgid, &templess_up, &templess_primary); | |
7c673cae FG |
1766 | if (real_primary == templess_primary){ |
1767 | ldout(cct, 10) << __func__ << " removing primary_temp " | |
1768 | << pgid << " -> " << real_primary | |
1769 | << " (unnecessary/redundant)" << dendl; | |
11fdf7f2 | 1770 | if (oldmap.primary_temp->count(pgid)) |
7c673cae FG |
1771 | pending_inc->new_primary_temp[pgid] = -1; |
1772 | else | |
1773 | pending_inc->new_primary_temp.erase(pgid); | |
1774 | } | |
1775 | } | |
1776 | } | |
1777 | ||
494da23a | 1778 | void OSDMap::get_upmap_pgs(vector<pg_t> *upmap_pgs) const |
94b18763 | 1779 | { |
494da23a TL |
1780 | upmap_pgs->reserve(pg_upmap.size() + pg_upmap_items.size()); |
1781 | for (auto& p : pg_upmap) | |
1782 | upmap_pgs->push_back(p.first); | |
1783 | for (auto& p : pg_upmap_items) | |
1784 | upmap_pgs->push_back(p.first); | |
1785 | } | |
94b18763 | 1786 | |
494da23a TL |
1787 | bool OSDMap::check_pg_upmaps( |
1788 | CephContext *cct, | |
1789 | const vector<pg_t>& to_check, | |
1790 | vector<pg_t> *to_cancel, | |
1791 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> *to_remap) const | |
1792 | { | |
1793 | bool any_change = false; | |
1794 | map<int, map<int, float>> rule_weight_map; | |
28e407b8 | 1795 | for (auto& pg : to_check) { |
494da23a | 1796 | const pg_pool_t *pi = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
1797 | if (!pi || pg.ps() >= pi->get_pg_num_pending()) { |
1798 | ldout(cct, 0) << __func__ << " pg " << pg << " is gone or merge source" | |
1799 | << dendl; | |
494da23a | 1800 | to_cancel->push_back(pg); |
11fdf7f2 TL |
1801 | continue; |
1802 | } | |
1803 | if (pi->is_pending_merge(pg, nullptr)) { | |
1804 | ldout(cct, 0) << __func__ << " pg " << pg << " is pending merge" | |
1805 | << dendl; | |
494da23a | 1806 | to_cancel->push_back(pg); |
94b18763 FG |
1807 | continue; |
1808 | } | |
494da23a TL |
1809 | vector<int> raw, up; |
1810 | pg_to_raw_upmap(pg, &raw, &up); | |
494da23a TL |
1811 | auto crush_rule = get_pg_pool_crush_rule(pg); |
1812 | auto r = crush->verify_upmap(cct, | |
1813 | crush_rule, | |
1814 | get_pg_pool_size(pg), | |
1815 | up); | |
a8e16298 TL |
1816 | if (r < 0) { |
1817 | ldout(cct, 0) << __func__ << " verify_upmap of pg " << pg | |
1818 | << " returning " << r | |
1819 | << dendl; | |
494da23a | 1820 | to_cancel->push_back(pg); |
a8e16298 TL |
1821 | continue; |
1822 | } | |
1823 | // below we check against crush-topology changing.. | |
28e407b8 AA |
1824 | map<int, float> weight_map; |
1825 | auto it = rule_weight_map.find(crush_rule); | |
1826 | if (it == rule_weight_map.end()) { | |
494da23a | 1827 | auto r = crush->get_rule_weight_osd_map(crush_rule, &weight_map); |
28e407b8 AA |
1828 | if (r < 0) { |
1829 | lderr(cct) << __func__ << " unable to get crush weight_map for " | |
494da23a TL |
1830 | << "crush_rule " << crush_rule |
1831 | << dendl; | |
28e407b8 AA |
1832 | continue; |
1833 | } | |
1834 | rule_weight_map[crush_rule] = weight_map; | |
1835 | } else { | |
1836 | weight_map = it->second; | |
1837 | } | |
28e407b8 | 1838 | ldout(cct, 10) << __func__ << " pg " << pg |
28e407b8 | 1839 | << " weight_map " << weight_map |
94b18763 | 1840 | << dendl; |
a8e16298 | 1841 | for (auto osd : up) { |
28e407b8 AA |
1842 | auto it = weight_map.find(osd); |
1843 | if (it == weight_map.end()) { | |
92f5a8d4 TL |
1844 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd << " is gone or has " |
1845 | << "been moved out of the specific crush-tree" | |
1846 | << dendl; | |
494da23a | 1847 | to_cancel->push_back(pg); |
94b18763 FG |
1848 | break; |
1849 | } | |
494da23a | 1850 | auto adjusted_weight = get_weightf(it->first) * it->second; |
28e407b8 | 1851 | if (adjusted_weight == 0) { |
92f5a8d4 TL |
1852 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd |
1853 | << " is out/crush-out" | |
1854 | << dendl; | |
494da23a | 1855 | to_cancel->push_back(pg); |
94b18763 FG |
1856 | break; |
1857 | } | |
1858 | } | |
eafe8130 TL |
1859 | if (!to_cancel->empty() && to_cancel->back() == pg) |
1860 | continue; | |
1861 | // okay, upmap is valid | |
1862 | // continue to check if it is still necessary | |
1863 | auto i = pg_upmap.find(pg); | |
1864 | if (i != pg_upmap.end() && raw == i->second) { | |
1865 | ldout(cct, 10) << " removing redundant pg_upmap " | |
1866 | << i->first << " " << i->second | |
1867 | << dendl; | |
1868 | to_cancel->push_back(pg); | |
1869 | continue; | |
1870 | } | |
1871 | auto j = pg_upmap_items.find(pg); | |
1872 | if (j != pg_upmap_items.end()) { | |
1873 | mempool::osdmap::vector<pair<int,int>> newmap; | |
1874 | for (auto& p : j->second) { | |
1875 | if (std::find(raw.begin(), raw.end(), p.first) == raw.end()) { | |
1876 | // cancel mapping if source osd does not exist anymore | |
1877 | continue; | |
1878 | } | |
1879 | if (p.second != CRUSH_ITEM_NONE && p.second < max_osd && | |
1880 | p.second >= 0 && osd_weight[p.second] == 0) { | |
1881 | // cancel mapping if target osd is out | |
1882 | continue; | |
1883 | } | |
1884 | newmap.push_back(p); | |
1885 | } | |
1886 | if (newmap.empty()) { | |
1887 | ldout(cct, 10) << " removing no-op pg_upmap_items " | |
1888 | << j->first << " " << j->second | |
1889 | << dendl; | |
1890 | to_cancel->push_back(pg); | |
1891 | } else if (newmap != j->second) { | |
1892 | ldout(cct, 10) << " simplifying partially no-op pg_upmap_items " | |
1893 | << j->first << " " << j->second | |
1894 | << " -> " << newmap | |
1895 | << dendl; | |
1896 | to_remap->insert({pg, newmap}); | |
1897 | any_change = true; | |
1898 | } | |
1899 | } | |
28e407b8 | 1900 | } |
494da23a TL |
1901 | any_change = any_change || !to_cancel->empty(); |
1902 | return any_change; | |
1903 | } | |
1904 | ||
1905 | void OSDMap::clean_pg_upmaps( | |
1906 | CephContext *cct, | |
1907 | Incremental *pending_inc, | |
1908 | const vector<pg_t>& to_cancel, | |
1909 | const map<pg_t, mempool::osdmap::vector<pair<int,int>>>& to_remap) const | |
1910 | { | |
28e407b8 | 1911 | for (auto &pg: to_cancel) { |
494da23a TL |
1912 | auto i = pending_inc->new_pg_upmap.find(pg); |
1913 | if (i != pending_inc->new_pg_upmap.end()) { | |
1914 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
1915 | << "pg_upmap entry " | |
1916 | << i->first << "->" << i->second | |
1917 | << dendl; | |
1918 | pending_inc->new_pg_upmap.erase(i); | |
94b18763 | 1919 | } |
494da23a TL |
1920 | auto j = pg_upmap.find(pg); |
1921 | if (j != pg_upmap.end()) { | |
1922 | ldout(cct, 10) << __func__ << " cancel invalid pg_upmap entry " | |
1923 | << j->first << "->" << j->second | |
1924 | << dendl; | |
1925 | pending_inc->old_pg_upmap.insert(pg); | |
1926 | } | |
1927 | auto p = pending_inc->new_pg_upmap_items.find(pg); | |
1928 | if (p != pending_inc->new_pg_upmap_items.end()) { | |
1929 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
1930 | << "pg_upmap_items entry " | |
1931 | << p->first << "->" << p->second | |
1932 | << dendl; | |
1933 | pending_inc->new_pg_upmap_items.erase(p); | |
1934 | } | |
1935 | auto q = pg_upmap_items.find(pg); | |
1936 | if (q != pg_upmap_items.end()) { | |
1937 | ldout(cct, 10) << __func__ << " cancel invalid " | |
1938 | << "pg_upmap_items entry " | |
1939 | << q->first << "->" << q->second | |
1940 | << dendl; | |
1941 | pending_inc->old_pg_upmap_items.insert(pg); | |
94b18763 FG |
1942 | } |
1943 | } | |
494da23a TL |
1944 | for (auto& i : to_remap) |
1945 | pending_inc->new_pg_upmap_items[i.first] = i.second; | |
1946 | } | |
1947 | ||
1948 | bool OSDMap::clean_pg_upmaps( | |
1949 | CephContext *cct, | |
1950 | Incremental *pending_inc) const | |
1951 | { | |
1952 | ldout(cct, 10) << __func__ << dendl; | |
1953 | vector<pg_t> to_check; | |
1954 | vector<pg_t> to_cancel; | |
1955 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> to_remap; | |
1956 | ||
1957 | get_upmap_pgs(&to_check); | |
1958 | auto any_change = check_pg_upmaps(cct, to_check, &to_cancel, &to_remap); | |
1959 | clean_pg_upmaps(cct, pending_inc, to_cancel, to_remap); | |
1960 | return any_change; | |
94b18763 FG |
1961 | } |
1962 | ||
7c673cae FG |
1963 | int OSDMap::apply_incremental(const Incremental &inc) |
1964 | { | |
1965 | new_blacklist_entries = false; | |
1966 | if (inc.epoch == 1) | |
1967 | fsid = inc.fsid; | |
1968 | else if (inc.fsid != fsid) | |
1969 | return -EINVAL; | |
1970 | ||
11fdf7f2 | 1971 | ceph_assert(inc.epoch == epoch+1); |
7c673cae FG |
1972 | |
1973 | epoch++; | |
1974 | modified = inc.modified; | |
1975 | ||
1976 | // full map? | |
1977 | if (inc.fullmap.length()) { | |
1978 | bufferlist bl(inc.fullmap); | |
1979 | decode(bl); | |
1980 | return 0; | |
1981 | } | |
1982 | ||
1983 | // nope, incremental. | |
31f18b77 | 1984 | if (inc.new_flags >= 0) { |
7c673cae | 1985 | flags = inc.new_flags; |
31f18b77 FG |
1986 | // the below is just to cover a newly-upgraded luminous mon |
1987 | // cluster that has to set require_jewel_osds or | |
1988 | // require_kraken_osds before the osds can be upgraded to | |
1989 | // luminous. | |
1990 | if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { | |
1991 | if (require_osd_release < CEPH_RELEASE_KRAKEN) { | |
1992 | require_osd_release = CEPH_RELEASE_KRAKEN; | |
1993 | } | |
1994 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
1995 | if (require_osd_release < CEPH_RELEASE_JEWEL) { | |
1996 | require_osd_release = CEPH_RELEASE_JEWEL; | |
1997 | } | |
1998 | } | |
1999 | } | |
7c673cae FG |
2000 | |
2001 | if (inc.new_max_osd >= 0) | |
2002 | set_max_osd(inc.new_max_osd); | |
2003 | ||
2004 | if (inc.new_pool_max != -1) | |
2005 | pool_max = inc.new_pool_max; | |
2006 | ||
2007 | for (const auto &pool : inc.new_pools) { | |
2008 | pools[pool.first] = pool.second; | |
2009 | pools[pool.first].last_change = epoch; | |
2010 | } | |
2011 | ||
11fdf7f2 TL |
2012 | new_removed_snaps = inc.new_removed_snaps; |
2013 | new_purged_snaps = inc.new_purged_snaps; | |
2014 | for (auto p = new_removed_snaps.begin(); | |
2015 | p != new_removed_snaps.end(); | |
2016 | ++p) { | |
2017 | removed_snaps_queue[p->first].union_of(p->second); | |
2018 | } | |
2019 | for (auto p = new_purged_snaps.begin(); | |
2020 | p != new_purged_snaps.end(); | |
2021 | ++p) { | |
2022 | auto q = removed_snaps_queue.find(p->first); | |
2023 | ceph_assert(q != removed_snaps_queue.end()); | |
2024 | q->second.subtract(p->second); | |
2025 | if (q->second.empty()) { | |
2026 | removed_snaps_queue.erase(q); | |
2027 | } | |
2028 | } | |
2029 | ||
2030 | if (inc.new_last_up_change != utime_t()) { | |
2031 | last_up_change = inc.new_last_up_change; | |
2032 | } | |
2033 | if (inc.new_last_in_change != utime_t()) { | |
2034 | last_in_change = inc.new_last_in_change; | |
2035 | } | |
2036 | ||
7c673cae FG |
2037 | for (const auto &pname : inc.new_pool_names) { |
2038 | auto pool_name_entry = pool_name.find(pname.first); | |
2039 | if (pool_name_entry != pool_name.end()) { | |
2040 | name_pool.erase(pool_name_entry->second); | |
2041 | pool_name_entry->second = pname.second; | |
2042 | } else { | |
2043 | pool_name[pname.first] = pname.second; | |
2044 | } | |
2045 | name_pool[pname.second] = pname.first; | |
2046 | } | |
2047 | ||
2048 | for (const auto &pool : inc.old_pools) { | |
2049 | pools.erase(pool); | |
2050 | name_pool.erase(pool_name[pool]); | |
2051 | pool_name.erase(pool); | |
2052 | } | |
2053 | ||
2054 | for (const auto &weight : inc.new_weight) { | |
2055 | set_weight(weight.first, weight.second); | |
2056 | ||
2057 | // if we are marking in, clear the AUTOOUT and NEW bits, and clear | |
2058 | // xinfo old_weight. | |
2059 | if (weight.second) { | |
2060 | osd_state[weight.first] &= ~(CEPH_OSD_AUTOOUT | CEPH_OSD_NEW); | |
2061 | osd_xinfo[weight.first].old_weight = 0; | |
2062 | } | |
2063 | } | |
2064 | ||
2065 | for (const auto &primary_affinity : inc.new_primary_affinity) { | |
2066 | set_primary_affinity(primary_affinity.first, primary_affinity.second); | |
2067 | } | |
2068 | ||
2069 | // erasure_code_profiles | |
2070 | for (const auto &profile : inc.old_erasure_code_profiles) | |
2071 | erasure_code_profiles.erase(profile); | |
2072 | ||
2073 | for (const auto &profile : inc.new_erasure_code_profiles) { | |
2074 | set_erasure_code_profile(profile.first, profile.second); | |
2075 | } | |
2076 | ||
2077 | // up/down | |
2078 | for (const auto &state : inc.new_state) { | |
2079 | const auto osd = state.first; | |
2080 | int s = state.second ? state.second : CEPH_OSD_UP; | |
2081 | if ((osd_state[osd] & CEPH_OSD_UP) && | |
2082 | (s & CEPH_OSD_UP)) { | |
2083 | osd_info[osd].down_at = epoch; | |
2084 | osd_xinfo[osd].down_stamp = modified; | |
2085 | } | |
2086 | if ((osd_state[osd] & CEPH_OSD_EXISTS) && | |
2087 | (s & CEPH_OSD_EXISTS)) { | |
2088 | // osd is destroyed; clear out anything interesting. | |
2089 | (*osd_uuid)[osd] = uuid_d(); | |
2090 | osd_info[osd] = osd_info_t(); | |
2091 | osd_xinfo[osd] = osd_xinfo_t(); | |
2092 | set_primary_affinity(osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
11fdf7f2 TL |
2093 | osd_addrs->client_addrs[osd].reset(new entity_addrvec_t()); |
2094 | osd_addrs->cluster_addrs[osd].reset(new entity_addrvec_t()); | |
2095 | osd_addrs->hb_front_addrs[osd].reset(new entity_addrvec_t()); | |
2096 | osd_addrs->hb_back_addrs[osd].reset(new entity_addrvec_t()); | |
7c673cae FG |
2097 | osd_state[osd] = 0; |
2098 | } else { | |
2099 | osd_state[osd] ^= s; | |
2100 | } | |
2101 | } | |
2102 | ||
2103 | for (const auto &client : inc.new_up_client) { | |
2104 | osd_state[client.first] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; | |
11fdf7f2 TL |
2105 | osd_addrs->client_addrs[client.first].reset( |
2106 | new entity_addrvec_t(client.second)); | |
2107 | osd_addrs->hb_back_addrs[client.first].reset( | |
2108 | new entity_addrvec_t(inc.new_hb_back_up.find(client.first)->second)); | |
2109 | osd_addrs->hb_front_addrs[client.first].reset( | |
2110 | new entity_addrvec_t(inc.new_hb_front_up.find(client.first)->second)); | |
7c673cae FG |
2111 | |
2112 | osd_info[client.first].up_from = epoch; | |
2113 | } | |
2114 | ||
2115 | for (const auto &cluster : inc.new_up_cluster) | |
11fdf7f2 TL |
2116 | osd_addrs->cluster_addrs[cluster.first].reset( |
2117 | new entity_addrvec_t(cluster.second)); | |
7c673cae FG |
2118 | |
2119 | // info | |
2120 | for (const auto &thru : inc.new_up_thru) | |
2121 | osd_info[thru.first].up_thru = thru.second; | |
2122 | ||
2123 | for (const auto &interval : inc.new_last_clean_interval) { | |
2124 | osd_info[interval.first].last_clean_begin = interval.second.first; | |
2125 | osd_info[interval.first].last_clean_end = interval.second.second; | |
2126 | } | |
2127 | ||
2128 | for (const auto &lost : inc.new_lost) | |
2129 | osd_info[lost.first].lost_at = lost.second; | |
2130 | ||
2131 | // xinfo | |
2132 | for (const auto &xinfo : inc.new_xinfo) | |
2133 | osd_xinfo[xinfo.first] = xinfo.second; | |
2134 | ||
2135 | // uuid | |
2136 | for (const auto &uuid : inc.new_uuid) | |
2137 | (*osd_uuid)[uuid.first] = uuid.second; | |
2138 | ||
2139 | // pg rebuild | |
2140 | for (const auto &pg : inc.new_pg_temp) { | |
2141 | if (pg.second.empty()) | |
2142 | pg_temp->erase(pg.first); | |
2143 | else | |
31f18b77 FG |
2144 | pg_temp->set(pg.first, pg.second); |
2145 | } | |
2146 | if (!inc.new_pg_temp.empty()) { | |
2147 | // make sure pg_temp is efficiently stored | |
2148 | pg_temp->rebuild(); | |
7c673cae FG |
2149 | } |
2150 | ||
2151 | for (const auto &pg : inc.new_primary_temp) { | |
2152 | if (pg.second == -1) | |
2153 | primary_temp->erase(pg.first); | |
2154 | else | |
2155 | (*primary_temp)[pg.first] = pg.second; | |
2156 | } | |
2157 | ||
2158 | for (auto& p : inc.new_pg_upmap) { | |
2159 | pg_upmap[p.first] = p.second; | |
2160 | } | |
2161 | for (auto& pg : inc.old_pg_upmap) { | |
2162 | pg_upmap.erase(pg); | |
2163 | } | |
2164 | for (auto& p : inc.new_pg_upmap_items) { | |
2165 | pg_upmap_items[p.first] = p.second; | |
2166 | } | |
2167 | for (auto& pg : inc.old_pg_upmap_items) { | |
2168 | pg_upmap_items.erase(pg); | |
2169 | } | |
2170 | ||
2171 | // blacklist | |
2172 | if (!inc.new_blacklist.empty()) { | |
2173 | blacklist.insert(inc.new_blacklist.begin(),inc.new_blacklist.end()); | |
2174 | new_blacklist_entries = true; | |
2175 | } | |
2176 | for (const auto &addr : inc.old_blacklist) | |
2177 | blacklist.erase(addr); | |
2178 | ||
81eedcae TL |
2179 | for (auto& i : inc.new_crush_node_flags) { |
2180 | if (i.second) { | |
2181 | crush_node_flags[i.first] = i.second; | |
2182 | } else { | |
2183 | crush_node_flags.erase(i.first); | |
2184 | } | |
2185 | } | |
2186 | ||
2187 | for (auto& i : inc.new_device_class_flags) { | |
2188 | if (i.second) { | |
2189 | device_class_flags[i.first] = i.second; | |
2190 | } else { | |
2191 | device_class_flags.erase(i.first); | |
2192 | } | |
2193 | } | |
2194 | ||
7c673cae FG |
2195 | // cluster snapshot? |
2196 | if (inc.cluster_snapshot.length()) { | |
2197 | cluster_snapshot = inc.cluster_snapshot; | |
2198 | cluster_snapshot_epoch = inc.epoch; | |
2199 | } else { | |
2200 | cluster_snapshot.clear(); | |
2201 | cluster_snapshot_epoch = 0; | |
2202 | } | |
2203 | ||
2204 | if (inc.new_nearfull_ratio >= 0) { | |
2205 | nearfull_ratio = inc.new_nearfull_ratio; | |
2206 | } | |
2207 | if (inc.new_backfillfull_ratio >= 0) { | |
2208 | backfillfull_ratio = inc.new_backfillfull_ratio; | |
2209 | } | |
2210 | if (inc.new_full_ratio >= 0) { | |
2211 | full_ratio = inc.new_full_ratio; | |
2212 | } | |
31f18b77 | 2213 | if (inc.new_require_min_compat_client > 0) { |
7c673cae FG |
2214 | require_min_compat_client = inc.new_require_min_compat_client; |
2215 | } | |
31f18b77 FG |
2216 | if (inc.new_require_osd_release >= 0) { |
2217 | require_osd_release = inc.new_require_osd_release; | |
2218 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { | |
2219 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 2220 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2221 | } |
2222 | } | |
7c673cae | 2223 | |
11fdf7f2 TL |
2224 | if (inc.new_require_osd_release >= 0) { |
2225 | require_osd_release = inc.new_require_osd_release; | |
2226 | if (require_osd_release >= CEPH_RELEASE_NAUTILUS) { | |
2227 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; | |
2228 | } | |
2229 | } | |
7c673cae FG |
2230 | // do new crush map last (after up/down stuff) |
2231 | if (inc.crush.length()) { | |
2232 | bufferlist bl(inc.crush); | |
11fdf7f2 | 2233 | auto blp = bl.cbegin(); |
7c673cae FG |
2234 | crush.reset(new CrushWrapper); |
2235 | crush->decode(blp); | |
31f18b77 FG |
2236 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { |
2237 | // only increment if this is a luminous-encoded osdmap, lest | |
2238 | // the mon's crush_version diverge from what the osds or others | |
2239 | // are decoding and applying on their end. if we won't encode | |
2240 | // it in the canonical version, don't change it. | |
2241 | ++crush_version; | |
2242 | } | |
81eedcae TL |
2243 | for (auto it = device_class_flags.begin(); |
2244 | it != device_class_flags.end();) { | |
2245 | const char* class_name = crush->get_class_name(it->first); | |
2246 | if (!class_name) // device class is gone | |
2247 | it = device_class_flags.erase(it); | |
2248 | else | |
2249 | it++; | |
2250 | } | |
7c673cae FG |
2251 | } |
2252 | ||
2253 | calc_num_osds(); | |
2254 | _calc_up_osd_features(); | |
2255 | return 0; | |
2256 | } | |
2257 | ||
2258 | // mapping | |
2259 | int OSDMap::map_to_pg( | |
2260 | int64_t poolid, | |
2261 | const string& name, | |
2262 | const string& key, | |
2263 | const string& nspace, | |
2264 | pg_t *pg) const | |
2265 | { | |
2266 | // calculate ps (placement seed) | |
2267 | const pg_pool_t *pool = get_pg_pool(poolid); | |
2268 | if (!pool) | |
2269 | return -ENOENT; | |
2270 | ps_t ps; | |
2271 | if (!key.empty()) | |
2272 | ps = pool->hash_key(key, nspace); | |
2273 | else | |
2274 | ps = pool->hash_key(name, nspace); | |
2275 | *pg = pg_t(ps, poolid); | |
2276 | return 0; | |
2277 | } | |
2278 | ||
2279 | int OSDMap::object_locator_to_pg( | |
2280 | const object_t& oid, const object_locator_t& loc, pg_t &pg) const | |
2281 | { | |
2282 | if (loc.hash >= 0) { | |
2283 | if (!get_pg_pool(loc.get_pool())) { | |
2284 | return -ENOENT; | |
2285 | } | |
2286 | pg = pg_t(loc.hash, loc.get_pool()); | |
2287 | return 0; | |
2288 | } | |
2289 | return map_to_pg(loc.get_pool(), oid.name, loc.key, loc.nspace, &pg); | |
2290 | } | |
2291 | ||
2292 | ceph_object_layout OSDMap::make_object_layout( | |
2293 | object_t oid, int pg_pool, string nspace) const | |
2294 | { | |
2295 | object_locator_t loc(pg_pool, nspace); | |
2296 | ||
2297 | ceph_object_layout ol; | |
2298 | pg_t pgid = object_locator_to_pg(oid, loc); | |
2299 | ol.ol_pgid = pgid.get_old_pg().v; | |
2300 | ol.ol_stripe_unit = 0; | |
2301 | return ol; | |
2302 | } | |
2303 | ||
2304 | void OSDMap::_remove_nonexistent_osds(const pg_pool_t& pool, | |
2305 | vector<int>& osds) const | |
2306 | { | |
2307 | if (pool.can_shift_osds()) { | |
2308 | unsigned removed = 0; | |
2309 | for (unsigned i = 0; i < osds.size(); i++) { | |
2310 | if (!exists(osds[i])) { | |
2311 | removed++; | |
2312 | continue; | |
2313 | } | |
2314 | if (removed) { | |
2315 | osds[i - removed] = osds[i]; | |
2316 | } | |
2317 | } | |
2318 | if (removed) | |
2319 | osds.resize(osds.size() - removed); | |
2320 | } else { | |
2321 | for (auto& osd : osds) { | |
2322 | if (!exists(osd)) | |
2323 | osd = CRUSH_ITEM_NONE; | |
2324 | } | |
2325 | } | |
2326 | } | |
2327 | ||
31f18b77 | 2328 | void OSDMap::_pg_to_raw_osds( |
7c673cae FG |
2329 | const pg_pool_t& pool, pg_t pg, |
2330 | vector<int> *osds, | |
2331 | ps_t *ppps) const | |
2332 | { | |
2333 | // map to osds[] | |
2334 | ps_t pps = pool.raw_pg_to_pps(pg); // placement ps | |
2335 | unsigned size = pool.get_size(); | |
2336 | ||
2337 | // what crush rule? | |
31f18b77 | 2338 | int ruleno = crush->find_rule(pool.get_crush_rule(), pool.get_type(), size); |
7c673cae FG |
2339 | if (ruleno >= 0) |
2340 | crush->do_rule(ruleno, pps, *osds, size, osd_weight, pg.pool()); | |
2341 | ||
2342 | _remove_nonexistent_osds(pool, *osds); | |
2343 | ||
2344 | if (ppps) | |
2345 | *ppps = pps; | |
7c673cae FG |
2346 | } |
2347 | ||
2348 | int OSDMap::_pick_primary(const vector<int>& osds) const | |
2349 | { | |
2350 | for (auto osd : osds) { | |
2351 | if (osd != CRUSH_ITEM_NONE) { | |
2352 | return osd; | |
2353 | } | |
2354 | } | |
2355 | return -1; | |
2356 | } | |
2357 | ||
224ce89b | 2358 | void OSDMap::_apply_upmap(const pg_pool_t& pi, pg_t raw_pg, vector<int> *raw) const |
7c673cae FG |
2359 | { |
2360 | pg_t pg = pi.raw_pg_to_pg(raw_pg); | |
2361 | auto p = pg_upmap.find(pg); | |
2362 | if (p != pg_upmap.end()) { | |
2363 | // make sure targets aren't marked out | |
2364 | for (auto osd : p->second) { | |
91327a77 AA |
2365 | if (osd != CRUSH_ITEM_NONE && osd < max_osd && osd >= 0 && |
2366 | osd_weight[osd] == 0) { | |
7c673cae FG |
2367 | // reject/ignore the explicit mapping |
2368 | return; | |
2369 | } | |
2370 | } | |
2371 | *raw = vector<int>(p->second.begin(), p->second.end()); | |
224ce89b | 2372 | // continue to check and apply pg_upmap_items if any |
7c673cae FG |
2373 | } |
2374 | ||
2375 | auto q = pg_upmap_items.find(pg); | |
2376 | if (q != pg_upmap_items.end()) { | |
181888fb FG |
2377 | // NOTE: this approach does not allow a bidirectional swap, |
2378 | // e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1]. | |
2379 | for (auto& r : q->second) { | |
2380 | // make sure the replacement value doesn't already appear | |
2381 | bool exists = false; | |
2382 | ssize_t pos = -1; | |
2383 | for (unsigned i = 0; i < raw->size(); ++i) { | |
2384 | int osd = (*raw)[i]; | |
2385 | if (osd == r.second) { | |
2386 | exists = true; | |
2387 | break; | |
2388 | } | |
2389 | // ignore mapping if target is marked out (or invalid osd id) | |
2390 | if (osd == r.first && | |
2391 | pos < 0 && | |
2392 | !(r.second != CRUSH_ITEM_NONE && r.second < max_osd && | |
91327a77 | 2393 | r.second >= 0 && osd_weight[r.second] == 0)) { |
181888fb FG |
2394 | pos = i; |
2395 | } | |
2396 | } | |
2397 | if (!exists && pos >= 0) { | |
2398 | (*raw)[pos] = r.second; | |
7c673cae FG |
2399 | } |
2400 | } | |
2401 | } | |
2402 | } | |
2403 | ||
2404 | // pg -> (up osd list) | |
2405 | void OSDMap::_raw_to_up_osds(const pg_pool_t& pool, const vector<int>& raw, | |
2406 | vector<int> *up) const | |
2407 | { | |
2408 | if (pool.can_shift_osds()) { | |
2409 | // shift left | |
2410 | up->clear(); | |
2411 | up->reserve(raw.size()); | |
2412 | for (unsigned i=0; i<raw.size(); i++) { | |
2413 | if (!exists(raw[i]) || is_down(raw[i])) | |
2414 | continue; | |
2415 | up->push_back(raw[i]); | |
2416 | } | |
2417 | } else { | |
2418 | // set down/dne devices to NONE | |
2419 | up->resize(raw.size()); | |
2420 | for (int i = raw.size() - 1; i >= 0; --i) { | |
2421 | if (!exists(raw[i]) || is_down(raw[i])) { | |
2422 | (*up)[i] = CRUSH_ITEM_NONE; | |
2423 | } else { | |
2424 | (*up)[i] = raw[i]; | |
2425 | } | |
2426 | } | |
2427 | } | |
2428 | } | |
2429 | ||
2430 | void OSDMap::_apply_primary_affinity(ps_t seed, | |
2431 | const pg_pool_t& pool, | |
2432 | vector<int> *osds, | |
2433 | int *primary) const | |
2434 | { | |
2435 | // do we have any non-default primary_affinity values for these osds? | |
2436 | if (!osd_primary_affinity) | |
2437 | return; | |
2438 | ||
2439 | bool any = false; | |
2440 | for (const auto osd : *osds) { | |
2441 | if (osd != CRUSH_ITEM_NONE && | |
2442 | (*osd_primary_affinity)[osd] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
2443 | any = true; | |
2444 | break; | |
2445 | } | |
2446 | } | |
2447 | if (!any) | |
2448 | return; | |
2449 | ||
2450 | // pick the primary. feed both the seed (for the pg) and the osd | |
2451 | // into the hash/rng so that a proportional fraction of an osd's pgs | |
2452 | // get rejected as primary. | |
2453 | int pos = -1; | |
2454 | for (unsigned i = 0; i < osds->size(); ++i) { | |
2455 | int o = (*osds)[i]; | |
2456 | if (o == CRUSH_ITEM_NONE) | |
2457 | continue; | |
2458 | unsigned a = (*osd_primary_affinity)[o]; | |
2459 | if (a < CEPH_OSD_MAX_PRIMARY_AFFINITY && | |
2460 | (crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
2461 | seed, o) >> 16) >= a) { | |
2462 | // we chose not to use this primary. note it anyway as a | |
2463 | // fallback in case we don't pick anyone else, but keep looking. | |
2464 | if (pos < 0) | |
2465 | pos = i; | |
2466 | } else { | |
2467 | pos = i; | |
2468 | break; | |
2469 | } | |
2470 | } | |
2471 | if (pos < 0) | |
2472 | return; | |
2473 | ||
2474 | *primary = (*osds)[pos]; | |
2475 | ||
2476 | if (pool.can_shift_osds() && pos > 0) { | |
2477 | // move the new primary to the front. | |
2478 | for (int i = pos; i > 0; --i) { | |
2479 | (*osds)[i] = (*osds)[i-1]; | |
2480 | } | |
2481 | (*osds)[0] = *primary; | |
2482 | } | |
2483 | } | |
2484 | ||
2485 | void OSDMap::_get_temp_osds(const pg_pool_t& pool, pg_t pg, | |
2486 | vector<int> *temp_pg, int *temp_primary) const | |
2487 | { | |
2488 | pg = pool.raw_pg_to_pg(pg); | |
2489 | const auto p = pg_temp->find(pg); | |
2490 | temp_pg->clear(); | |
2491 | if (p != pg_temp->end()) { | |
2492 | for (unsigned i=0; i<p->second.size(); i++) { | |
2493 | if (!exists(p->second[i]) || is_down(p->second[i])) { | |
2494 | if (pool.can_shift_osds()) { | |
2495 | continue; | |
2496 | } else { | |
2497 | temp_pg->push_back(CRUSH_ITEM_NONE); | |
2498 | } | |
2499 | } else { | |
2500 | temp_pg->push_back(p->second[i]); | |
2501 | } | |
2502 | } | |
2503 | } | |
2504 | const auto &pp = primary_temp->find(pg); | |
2505 | *temp_primary = -1; | |
2506 | if (pp != primary_temp->end()) { | |
2507 | *temp_primary = pp->second; | |
2508 | } else if (!temp_pg->empty()) { // apply pg_temp's primary | |
2509 | for (unsigned i = 0; i < temp_pg->size(); ++i) { | |
2510 | if ((*temp_pg)[i] != CRUSH_ITEM_NONE) { | |
2511 | *temp_primary = (*temp_pg)[i]; | |
2512 | break; | |
2513 | } | |
2514 | } | |
2515 | } | |
2516 | } | |
2517 | ||
31f18b77 | 2518 | void OSDMap::pg_to_raw_osds(pg_t pg, vector<int> *raw, int *primary) const |
7c673cae | 2519 | { |
7c673cae | 2520 | const pg_pool_t *pool = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2521 | if (!pool) { |
2522 | *primary = -1; | |
2523 | raw->clear(); | |
31f18b77 | 2524 | return; |
11fdf7f2 | 2525 | } |
31f18b77 | 2526 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
11fdf7f2 | 2527 | *primary = _pick_primary(*raw); |
7c673cae FG |
2528 | } |
2529 | ||
494da23a TL |
2530 | void OSDMap::pg_to_raw_upmap(pg_t pg, vector<int>*raw, |
2531 | vector<int> *raw_upmap) const | |
a8e16298 TL |
2532 | { |
2533 | auto pool = get_pg_pool(pg.pool()); | |
2534 | if (!pool) { | |
2535 | raw_upmap->clear(); | |
2536 | return; | |
2537 | } | |
494da23a TL |
2538 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
2539 | *raw_upmap = *raw; | |
a8e16298 TL |
2540 | _apply_upmap(*pool, pg, raw_upmap); |
2541 | } | |
2542 | ||
7c673cae FG |
2543 | void OSDMap::pg_to_raw_up(pg_t pg, vector<int> *up, int *primary) const |
2544 | { | |
2545 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2546 | if (!pool) { | |
11fdf7f2 TL |
2547 | *primary = -1; |
2548 | up->clear(); | |
7c673cae FG |
2549 | return; |
2550 | } | |
2551 | vector<int> raw; | |
2552 | ps_t pps; | |
2553 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2554 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2555 | _raw_to_up_osds(*pool, raw, up); |
2556 | *primary = _pick_primary(raw); | |
2557 | _apply_primary_affinity(pps, *pool, up, primary); | |
2558 | } | |
31f18b77 | 2559 | |
7c673cae FG |
2560 | void OSDMap::_pg_to_up_acting_osds( |
2561 | const pg_t& pg, vector<int> *up, int *up_primary, | |
2562 | vector<int> *acting, int *acting_primary, | |
2563 | bool raw_pg_to_pg) const | |
2564 | { | |
2565 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2566 | if (!pool || | |
2567 | (!raw_pg_to_pg && pg.ps() >= pool->get_pg_num())) { | |
2568 | if (up) | |
2569 | up->clear(); | |
2570 | if (up_primary) | |
2571 | *up_primary = -1; | |
2572 | if (acting) | |
2573 | acting->clear(); | |
2574 | if (acting_primary) | |
2575 | *acting_primary = -1; | |
2576 | return; | |
2577 | } | |
2578 | vector<int> raw; | |
2579 | vector<int> _up; | |
2580 | vector<int> _acting; | |
2581 | int _up_primary; | |
2582 | int _acting_primary; | |
2583 | ps_t pps; | |
2584 | _get_temp_osds(*pool, pg, &_acting, &_acting_primary); | |
2585 | if (_acting.empty() || up || up_primary) { | |
2586 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2587 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2588 | _raw_to_up_osds(*pool, raw, &_up); |
2589 | _up_primary = _pick_primary(_up); | |
2590 | _apply_primary_affinity(pps, *pool, &_up, &_up_primary); | |
2591 | if (_acting.empty()) { | |
2592 | _acting = _up; | |
2593 | if (_acting_primary == -1) { | |
2594 | _acting_primary = _up_primary; | |
2595 | } | |
2596 | } | |
2597 | ||
2598 | if (up) | |
2599 | up->swap(_up); | |
2600 | if (up_primary) | |
2601 | *up_primary = _up_primary; | |
2602 | } | |
2603 | ||
2604 | if (acting) | |
2605 | acting->swap(_acting); | |
2606 | if (acting_primary) | |
2607 | *acting_primary = _acting_primary; | |
2608 | } | |
2609 | ||
2610 | int OSDMap::calc_pg_rank(int osd, const vector<int>& acting, int nrep) | |
2611 | { | |
2612 | if (!nrep) | |
2613 | nrep = acting.size(); | |
2614 | for (int i=0; i<nrep; i++) | |
2615 | if (acting[i] == osd) | |
2616 | return i; | |
2617 | return -1; | |
2618 | } | |
2619 | ||
2620 | int OSDMap::calc_pg_role(int osd, const vector<int>& acting, int nrep) | |
2621 | { | |
2622 | return calc_pg_rank(osd, acting, nrep); | |
2623 | } | |
2624 | ||
2625 | bool OSDMap::primary_changed( | |
2626 | int oldprimary, | |
2627 | const vector<int> &oldacting, | |
2628 | int newprimary, | |
2629 | const vector<int> &newacting) | |
2630 | { | |
2631 | if (oldacting.empty() && newacting.empty()) | |
2632 | return false; // both still empty | |
2633 | if (oldacting.empty() ^ newacting.empty()) | |
2634 | return true; // was empty, now not, or vice versa | |
2635 | if (oldprimary != newprimary) | |
2636 | return true; // primary changed | |
2637 | if (calc_pg_rank(oldprimary, oldacting) != | |
2638 | calc_pg_rank(newprimary, newacting)) | |
2639 | return true; | |
2640 | return false; // same primary (tho replicas may have changed) | |
2641 | } | |
2642 | ||
28e407b8 AA |
2643 | uint64_t OSDMap::get_encoding_features() const |
2644 | { | |
2645 | uint64_t f = SIGNIFICANT_FEATURES; | |
11fdf7f2 TL |
2646 | if (require_osd_release < CEPH_RELEASE_NAUTILUS) { |
2647 | f &= ~CEPH_FEATURE_SERVER_NAUTILUS; | |
2648 | } | |
2649 | if (require_osd_release < CEPH_RELEASE_MIMIC) { | |
2650 | f &= ~CEPH_FEATURE_SERVER_MIMIC; | |
2651 | } | |
28e407b8 AA |
2652 | if (require_osd_release < CEPH_RELEASE_LUMINOUS) { |
2653 | f &= ~(CEPH_FEATURE_SERVER_LUMINOUS | | |
2654 | CEPH_FEATURE_CRUSH_CHOOSE_ARGS); | |
2655 | } | |
2656 | if (require_osd_release < CEPH_RELEASE_KRAKEN) { | |
2657 | f &= ~(CEPH_FEATURE_SERVER_KRAKEN | | |
1adf2230 | 2658 | CEPH_FEATURE_MSG_ADDR2); |
28e407b8 AA |
2659 | } |
2660 | if (require_osd_release < CEPH_RELEASE_JEWEL) { | |
2661 | f &= ~(CEPH_FEATURE_SERVER_JEWEL | | |
1adf2230 AA |
2662 | CEPH_FEATURE_NEW_OSDOP_ENCODING | |
2663 | CEPH_FEATURE_CRUSH_TUNABLES5); | |
28e407b8 AA |
2664 | } |
2665 | return f; | |
2666 | } | |
7c673cae FG |
2667 | |
2668 | // serialize, unserialize | |
2669 | void OSDMap::encode_client_old(bufferlist& bl) const | |
2670 | { | |
11fdf7f2 | 2671 | using ceph::encode; |
7c673cae | 2672 | __u16 v = 5; |
11fdf7f2 | 2673 | encode(v, bl); |
7c673cae FG |
2674 | |
2675 | // base | |
11fdf7f2 TL |
2676 | encode(fsid, bl); |
2677 | encode(epoch, bl); | |
2678 | encode(created, bl); | |
2679 | encode(modified, bl); | |
7c673cae | 2680 | |
11fdf7f2 | 2681 | // for encode(pools, bl); |
7c673cae | 2682 | __u32 n = pools.size(); |
11fdf7f2 | 2683 | encode(n, bl); |
7c673cae FG |
2684 | |
2685 | for (const auto &pool : pools) { | |
2686 | n = pool.first; | |
11fdf7f2 TL |
2687 | encode(n, bl); |
2688 | encode(pool.second, bl, 0); | |
7c673cae | 2689 | } |
11fdf7f2 | 2690 | // for encode(pool_name, bl); |
7c673cae | 2691 | n = pool_name.size(); |
11fdf7f2 | 2692 | encode(n, bl); |
7c673cae FG |
2693 | for (const auto &pname : pool_name) { |
2694 | n = pname.first; | |
11fdf7f2 TL |
2695 | encode(n, bl); |
2696 | encode(pname.second, bl); | |
7c673cae | 2697 | } |
11fdf7f2 | 2698 | // for encode(pool_max, bl); |
7c673cae | 2699 | n = pool_max; |
11fdf7f2 | 2700 | encode(n, bl); |
7c673cae | 2701 | |
11fdf7f2 | 2702 | encode(flags, bl); |
7c673cae | 2703 | |
11fdf7f2 | 2704 | encode(max_osd, bl); |
31f18b77 FG |
2705 | { |
2706 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2707 | encode(n, bl); |
31f18b77 | 2708 | for (auto s : osd_state) { |
11fdf7f2 | 2709 | encode((uint8_t)s, bl); |
31f18b77 FG |
2710 | } |
2711 | } | |
11fdf7f2 TL |
2712 | encode(osd_weight, bl); |
2713 | encode(osd_addrs->client_addrs, bl, 0); | |
7c673cae | 2714 | |
11fdf7f2 | 2715 | // for encode(pg_temp, bl); |
7c673cae | 2716 | n = pg_temp->size(); |
11fdf7f2 | 2717 | encode(n, bl); |
7c673cae FG |
2718 | for (const auto pg : *pg_temp) { |
2719 | old_pg_t opg = pg.first.get_old_pg(); | |
11fdf7f2 TL |
2720 | encode(opg, bl); |
2721 | encode(pg.second, bl); | |
7c673cae FG |
2722 | } |
2723 | ||
2724 | // crush | |
2725 | bufferlist cbl; | |
2726 | crush->encode(cbl, 0 /* legacy (no) features */); | |
11fdf7f2 | 2727 | encode(cbl, bl); |
7c673cae FG |
2728 | } |
2729 | ||
2730 | void OSDMap::encode_classic(bufferlist& bl, uint64_t features) const | |
2731 | { | |
11fdf7f2 | 2732 | using ceph::encode; |
7c673cae FG |
2733 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
2734 | encode_client_old(bl); | |
2735 | return; | |
2736 | } | |
2737 | ||
2738 | __u16 v = 6; | |
11fdf7f2 | 2739 | encode(v, bl); |
7c673cae FG |
2740 | |
2741 | // base | |
11fdf7f2 TL |
2742 | encode(fsid, bl); |
2743 | encode(epoch, bl); | |
2744 | encode(created, bl); | |
2745 | encode(modified, bl); | |
7c673cae | 2746 | |
11fdf7f2 TL |
2747 | encode(pools, bl, features); |
2748 | encode(pool_name, bl); | |
2749 | encode(pool_max, bl); | |
7c673cae | 2750 | |
11fdf7f2 | 2751 | encode(flags, bl); |
7c673cae | 2752 | |
11fdf7f2 | 2753 | encode(max_osd, bl); |
31f18b77 FG |
2754 | { |
2755 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2756 | encode(n, bl); |
31f18b77 | 2757 | for (auto s : osd_state) { |
11fdf7f2 | 2758 | encode((uint8_t)s, bl); |
31f18b77 FG |
2759 | } |
2760 | } | |
11fdf7f2 TL |
2761 | encode(osd_weight, bl); |
2762 | encode(osd_addrs->client_addrs, bl, features); | |
7c673cae | 2763 | |
11fdf7f2 | 2764 | encode(*pg_temp, bl); |
7c673cae FG |
2765 | |
2766 | // crush | |
2767 | bufferlist cbl; | |
2768 | crush->encode(cbl, 0 /* legacy (no) features */); | |
11fdf7f2 | 2769 | encode(cbl, bl); |
7c673cae FG |
2770 | |
2771 | // extended | |
2772 | __u16 ev = 10; | |
11fdf7f2 TL |
2773 | encode(ev, bl); |
2774 | encode(osd_addrs->hb_back_addrs, bl, features); | |
2775 | encode(osd_info, bl); | |
2776 | encode(blacklist, bl, features); | |
2777 | encode(osd_addrs->cluster_addrs, bl, features); | |
2778 | encode(cluster_snapshot_epoch, bl); | |
2779 | encode(cluster_snapshot, bl); | |
2780 | encode(*osd_uuid, bl); | |
2781 | encode(osd_xinfo, bl); | |
2782 | encode(osd_addrs->hb_front_addrs, bl, features); | |
7c673cae FG |
2783 | } |
2784 | ||
11fdf7f2 TL |
2785 | /* for a description of osdmap versions, and when they were introduced, please |
2786 | * refer to | |
2787 | * doc/dev/osd_internals/osdmap_versions.txt | |
2788 | */ | |
7c673cae FG |
2789 | void OSDMap::encode(bufferlist& bl, uint64_t features) const |
2790 | { | |
11fdf7f2 | 2791 | using ceph::encode; |
7c673cae FG |
2792 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
2793 | encode_classic(bl, features); | |
2794 | return; | |
2795 | } | |
2796 | ||
2797 | // only a select set of callers should *ever* be encoding new | |
2798 | // OSDMaps. others should be passing around the canonical encoded | |
2799 | // buffers from on high. select out those callers by passing in an | |
2800 | // "impossible" feature bit. | |
11fdf7f2 | 2801 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
2802 | features &= ~CEPH_FEATURE_RESERVED; |
2803 | ||
2804 | size_t start_offset = bl.length(); | |
2805 | size_t tail_offset; | |
11fdf7f2 TL |
2806 | size_t crc_offset; |
2807 | std::optional<buffer::list::contiguous_filler> crc_filler; | |
7c673cae FG |
2808 | |
2809 | // meta-encoding: how we include client-used and osd-specific data | |
2810 | ENCODE_START(8, 7, bl); | |
2811 | ||
2812 | { | |
28e407b8 AA |
2813 | // NOTE: any new encoding dependencies must be reflected by |
2814 | // SIGNIFICANT_FEATURES | |
11fdf7f2 | 2815 | uint8_t v = 9; |
31f18b77 | 2816 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
7c673cae | 2817 | v = 3; |
11fdf7f2 TL |
2818 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
2819 | v = 6; | |
2820 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
2821 | v = 7; | |
7c673cae FG |
2822 | } |
2823 | ENCODE_START(v, 1, bl); // client-usable data | |
2824 | // base | |
11fdf7f2 TL |
2825 | encode(fsid, bl); |
2826 | encode(epoch, bl); | |
2827 | encode(created, bl); | |
2828 | encode(modified, bl); | |
7c673cae | 2829 | |
11fdf7f2 TL |
2830 | encode(pools, bl, features); |
2831 | encode(pool_name, bl); | |
2832 | encode(pool_max, bl); | |
7c673cae | 2833 | |
31f18b77 FG |
2834 | if (v < 4) { |
2835 | decltype(flags) f = flags; | |
2836 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) | |
c07f9fc5 | 2837 | f |= CEPH_OSDMAP_REQUIRE_LUMINOUS | CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2838 | else if (require_osd_release == CEPH_RELEASE_KRAKEN) |
2839 | f |= CEPH_OSDMAP_REQUIRE_KRAKEN; | |
2840 | else if (require_osd_release == CEPH_RELEASE_JEWEL) | |
2841 | f |= CEPH_OSDMAP_REQUIRE_JEWEL; | |
11fdf7f2 | 2842 | encode(f, bl); |
31f18b77 | 2843 | } else { |
11fdf7f2 | 2844 | encode(flags, bl); |
31f18b77 | 2845 | } |
7c673cae | 2846 | |
11fdf7f2 | 2847 | encode(max_osd, bl); |
31f18b77 | 2848 | if (v >= 5) { |
11fdf7f2 | 2849 | encode(osd_state, bl); |
31f18b77 FG |
2850 | } else { |
2851 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2852 | encode(n, bl); |
31f18b77 | 2853 | for (auto s : osd_state) { |
11fdf7f2 | 2854 | encode((uint8_t)s, bl); |
31f18b77 FG |
2855 | } |
2856 | } | |
11fdf7f2 TL |
2857 | encode(osd_weight, bl); |
2858 | if (v >= 8) { | |
2859 | encode(osd_addrs->client_addrs, bl, features); | |
2860 | } else { | |
2861 | encode_addrvec_pvec_as_addr(osd_addrs->client_addrs, bl, features); | |
2862 | } | |
7c673cae | 2863 | |
11fdf7f2 TL |
2864 | encode(*pg_temp, bl); |
2865 | encode(*primary_temp, bl); | |
7c673cae | 2866 | if (osd_primary_affinity) { |
11fdf7f2 | 2867 | encode(*osd_primary_affinity, bl); |
7c673cae FG |
2868 | } else { |
2869 | vector<__u32> v; | |
11fdf7f2 | 2870 | encode(v, bl); |
7c673cae FG |
2871 | } |
2872 | ||
2873 | // crush | |
2874 | bufferlist cbl; | |
2875 | crush->encode(cbl, features); | |
11fdf7f2 TL |
2876 | encode(cbl, bl); |
2877 | encode(erasure_code_profiles, bl); | |
7c673cae FG |
2878 | |
2879 | if (v >= 4) { | |
11fdf7f2 TL |
2880 | encode(pg_upmap, bl); |
2881 | encode(pg_upmap_items, bl); | |
7c673cae | 2882 | } else { |
11fdf7f2 TL |
2883 | ceph_assert(pg_upmap.empty()); |
2884 | ceph_assert(pg_upmap_items.empty()); | |
7c673cae | 2885 | } |
31f18b77 | 2886 | if (v >= 6) { |
11fdf7f2 TL |
2887 | encode(crush_version, bl); |
2888 | } | |
2889 | if (v >= 7) { | |
2890 | encode(new_removed_snaps, bl); | |
2891 | encode(new_purged_snaps, bl); | |
2892 | } | |
2893 | if (v >= 9) { | |
2894 | encode(last_up_change, bl); | |
2895 | encode(last_in_change, bl); | |
31f18b77 | 2896 | } |
7c673cae FG |
2897 | ENCODE_FINISH(bl); // client-usable data |
2898 | } | |
2899 | ||
2900 | { | |
28e407b8 AA |
2901 | // NOTE: any new encoding dependencies must be reflected by |
2902 | // SIGNIFICANT_FEATURES | |
81eedcae | 2903 | uint8_t target_v = 9; |
7c673cae FG |
2904 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
2905 | target_v = 1; | |
11fdf7f2 TL |
2906 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
2907 | target_v = 5; | |
2908 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
2909 | target_v = 6; | |
7c673cae FG |
2910 | } |
2911 | ENCODE_START(target_v, 1, bl); // extended, osd-only data | |
11fdf7f2 TL |
2912 | if (target_v < 7) { |
2913 | encode_addrvec_pvec_as_addr(osd_addrs->hb_back_addrs, bl, features); | |
2914 | } else { | |
2915 | encode(osd_addrs->hb_back_addrs, bl, features); | |
2916 | } | |
2917 | encode(osd_info, bl); | |
7c673cae FG |
2918 | { |
2919 | // put this in a sorted, ordered map<> so that we encode in a | |
2920 | // deterministic order. | |
2921 | map<entity_addr_t,utime_t> blacklist_map; | |
2922 | for (const auto &addr : blacklist) | |
2923 | blacklist_map.insert(make_pair(addr.first, addr.second)); | |
11fdf7f2 TL |
2924 | encode(blacklist_map, bl, features); |
2925 | } | |
2926 | if (target_v < 7) { | |
2927 | encode_addrvec_pvec_as_addr(osd_addrs->cluster_addrs, bl, features); | |
2928 | } else { | |
2929 | encode(osd_addrs->cluster_addrs, bl, features); | |
2930 | } | |
2931 | encode(cluster_snapshot_epoch, bl); | |
2932 | encode(cluster_snapshot, bl); | |
2933 | encode(*osd_uuid, bl); | |
2934 | encode(osd_xinfo, bl); | |
2935 | if (target_v < 7) { | |
2936 | encode_addrvec_pvec_as_addr(osd_addrs->hb_front_addrs, bl, features); | |
2937 | } else { | |
2938 | encode(osd_addrs->hb_front_addrs, bl, features); | |
2939 | } | |
7c673cae | 2940 | if (target_v >= 2) { |
11fdf7f2 TL |
2941 | encode(nearfull_ratio, bl); |
2942 | encode(full_ratio, bl); | |
2943 | encode(backfillfull_ratio, bl); | |
31f18b77 FG |
2944 | } |
2945 | // 4 was string-based new_require_min_compat_client | |
2946 | if (target_v >= 5) { | |
11fdf7f2 TL |
2947 | encode(require_min_compat_client, bl); |
2948 | encode(require_osd_release, bl); | |
2949 | } | |
2950 | if (target_v >= 6) { | |
2951 | encode(removed_snaps_queue, bl); | |
7c673cae | 2952 | } |
81eedcae TL |
2953 | if (target_v >= 8) { |
2954 | encode(crush_node_flags, bl); | |
2955 | } | |
2956 | if (target_v >= 9) { | |
2957 | encode(device_class_flags, bl); | |
2958 | } | |
7c673cae FG |
2959 | ENCODE_FINISH(bl); // osd-only data |
2960 | } | |
2961 | ||
11fdf7f2 TL |
2962 | crc_offset = bl.length(); |
2963 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
2964 | tail_offset = bl.length(); |
2965 | ||
2966 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
2967 | ||
2968 | // fill in crc | |
2969 | bufferlist front; | |
11fdf7f2 | 2970 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae FG |
2971 | crc = front.crc32c(-1); |
2972 | if (tail_offset < bl.length()) { | |
2973 | bufferlist tail; | |
2974 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); | |
2975 | crc = tail.crc32c(crc); | |
2976 | } | |
2977 | ceph_le32 crc_le; | |
2978 | crc_le = crc; | |
11fdf7f2 | 2979 | crc_filler->copy_in(4, (char*)&crc_le); |
7c673cae FG |
2980 | crc_defined = true; |
2981 | } | |
2982 | ||
11fdf7f2 TL |
2983 | /* for a description of osdmap versions, and when they were introduced, please |
2984 | * refer to | |
2985 | * doc/dev/osd_internals/osdmap_versions.txt | |
2986 | */ | |
7c673cae FG |
2987 | void OSDMap::decode(bufferlist& bl) |
2988 | { | |
11fdf7f2 | 2989 | auto p = bl.cbegin(); |
7c673cae FG |
2990 | decode(p); |
2991 | } | |
2992 | ||
11fdf7f2 | 2993 | void OSDMap::decode_classic(bufferlist::const_iterator& p) |
7c673cae | 2994 | { |
11fdf7f2 | 2995 | using ceph::decode; |
7c673cae FG |
2996 | __u32 n, t; |
2997 | __u16 v; | |
11fdf7f2 | 2998 | decode(v, p); |
7c673cae FG |
2999 | |
3000 | // base | |
11fdf7f2 TL |
3001 | decode(fsid, p); |
3002 | decode(epoch, p); | |
3003 | decode(created, p); | |
3004 | decode(modified, p); | |
7c673cae FG |
3005 | |
3006 | if (v < 6) { | |
3007 | if (v < 4) { | |
3008 | int32_t max_pools = 0; | |
11fdf7f2 | 3009 | decode(max_pools, p); |
7c673cae FG |
3010 | pool_max = max_pools; |
3011 | } | |
3012 | pools.clear(); | |
11fdf7f2 | 3013 | decode(n, p); |
7c673cae | 3014 | while (n--) { |
11fdf7f2 TL |
3015 | decode(t, p); |
3016 | decode(pools[t], p); | |
7c673cae FG |
3017 | } |
3018 | if (v == 4) { | |
11fdf7f2 | 3019 | decode(n, p); |
7c673cae FG |
3020 | pool_max = n; |
3021 | } else if (v == 5) { | |
3022 | pool_name.clear(); | |
11fdf7f2 | 3023 | decode(n, p); |
7c673cae | 3024 | while (n--) { |
11fdf7f2 TL |
3025 | decode(t, p); |
3026 | decode(pool_name[t], p); | |
7c673cae | 3027 | } |
11fdf7f2 | 3028 | decode(n, p); |
7c673cae FG |
3029 | pool_max = n; |
3030 | } | |
3031 | } else { | |
11fdf7f2 TL |
3032 | decode(pools, p); |
3033 | decode(pool_name, p); | |
3034 | decode(pool_max, p); | |
7c673cae FG |
3035 | } |
3036 | // kludge around some old bug that zeroed out pool_max (#2307) | |
3037 | if (pools.size() && pool_max < pools.rbegin()->first) { | |
3038 | pool_max = pools.rbegin()->first; | |
3039 | } | |
3040 | ||
11fdf7f2 | 3041 | decode(flags, p); |
7c673cae | 3042 | |
11fdf7f2 | 3043 | decode(max_osd, p); |
31f18b77 FG |
3044 | { |
3045 | vector<uint8_t> os; | |
11fdf7f2 | 3046 | decode(os, p); |
31f18b77 FG |
3047 | osd_state.resize(os.size()); |
3048 | for (unsigned i = 0; i < os.size(); ++i) { | |
3049 | osd_state[i] = os[i]; | |
3050 | } | |
3051 | } | |
11fdf7f2 TL |
3052 | decode(osd_weight, p); |
3053 | decode(osd_addrs->client_addrs, p); | |
7c673cae FG |
3054 | if (v <= 5) { |
3055 | pg_temp->clear(); | |
11fdf7f2 | 3056 | decode(n, p); |
7c673cae FG |
3057 | while (n--) { |
3058 | old_pg_t opg; | |
3059 | ::decode_raw(opg, p); | |
31f18b77 | 3060 | mempool::osdmap::vector<int32_t> v; |
11fdf7f2 | 3061 | decode(v, p); |
31f18b77 | 3062 | pg_temp->set(pg_t(opg), v); |
7c673cae FG |
3063 | } |
3064 | } else { | |
11fdf7f2 | 3065 | decode(*pg_temp, p); |
7c673cae FG |
3066 | } |
3067 | ||
3068 | // crush | |
3069 | bufferlist cbl; | |
11fdf7f2 TL |
3070 | decode(cbl, p); |
3071 | auto cblp = cbl.cbegin(); | |
7c673cae FG |
3072 | crush->decode(cblp); |
3073 | ||
3074 | // extended | |
3075 | __u16 ev = 0; | |
3076 | if (v >= 5) | |
11fdf7f2 TL |
3077 | decode(ev, p); |
3078 | decode(osd_addrs->hb_back_addrs, p); | |
3079 | decode(osd_info, p); | |
7c673cae | 3080 | if (v < 5) |
11fdf7f2 | 3081 | decode(pool_name, p); |
7c673cae | 3082 | |
11fdf7f2 | 3083 | decode(blacklist, p); |
7c673cae | 3084 | if (ev >= 6) |
11fdf7f2 | 3085 | decode(osd_addrs->cluster_addrs, p); |
7c673cae | 3086 | else |
11fdf7f2 | 3087 | osd_addrs->cluster_addrs.resize(osd_addrs->client_addrs.size()); |
7c673cae FG |
3088 | |
3089 | if (ev >= 7) { | |
11fdf7f2 TL |
3090 | decode(cluster_snapshot_epoch, p); |
3091 | decode(cluster_snapshot, p); | |
7c673cae FG |
3092 | } |
3093 | ||
3094 | if (ev >= 8) { | |
11fdf7f2 | 3095 | decode(*osd_uuid, p); |
7c673cae FG |
3096 | } else { |
3097 | osd_uuid->resize(max_osd); | |
3098 | } | |
3099 | if (ev >= 9) | |
11fdf7f2 | 3100 | decode(osd_xinfo, p); |
7c673cae FG |
3101 | else |
3102 | osd_xinfo.resize(max_osd); | |
3103 | ||
3104 | if (ev >= 10) | |
11fdf7f2 | 3105 | decode(osd_addrs->hb_front_addrs, p); |
7c673cae | 3106 | else |
11fdf7f2 | 3107 | osd_addrs->hb_front_addrs.resize(osd_addrs->hb_back_addrs.size()); |
7c673cae FG |
3108 | |
3109 | osd_primary_affinity.reset(); | |
3110 | ||
3111 | post_decode(); | |
3112 | } | |
3113 | ||
11fdf7f2 | 3114 | void OSDMap::decode(bufferlist::const_iterator& bl) |
7c673cae | 3115 | { |
11fdf7f2 | 3116 | using ceph::decode; |
7c673cae FG |
3117 | /** |
3118 | * Older encodings of the OSDMap had a single struct_v which | |
3119 | * covered the whole encoding, and was prior to our modern | |
3120 | * stuff which includes a compatv and a size. So if we see | |
3121 | * a struct_v < 7, we must rewind to the beginning and use our | |
3122 | * classic decoder. | |
3123 | */ | |
3124 | size_t start_offset = bl.get_off(); | |
3125 | size_t tail_offset = 0; | |
3126 | bufferlist crc_front, crc_tail; | |
3127 | ||
3128 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
3129 | if (struct_v < 7) { | |
11fdf7f2 | 3130 | bl.seek(start_offset); |
7c673cae FG |
3131 | decode_classic(bl); |
3132 | return; | |
3133 | } | |
3134 | /** | |
3135 | * Since we made it past that hurdle, we can use our normal paths. | |
3136 | */ | |
3137 | { | |
11fdf7f2 | 3138 | DECODE_START(9, bl); // client-usable data |
7c673cae | 3139 | // base |
11fdf7f2 TL |
3140 | decode(fsid, bl); |
3141 | decode(epoch, bl); | |
3142 | decode(created, bl); | |
3143 | decode(modified, bl); | |
7c673cae | 3144 | |
11fdf7f2 TL |
3145 | decode(pools, bl); |
3146 | decode(pool_name, bl); | |
3147 | decode(pool_max, bl); | |
7c673cae | 3148 | |
11fdf7f2 | 3149 | decode(flags, bl); |
7c673cae | 3150 | |
11fdf7f2 | 3151 | decode(max_osd, bl); |
31f18b77 | 3152 | if (struct_v >= 5) { |
11fdf7f2 | 3153 | decode(osd_state, bl); |
31f18b77 FG |
3154 | } else { |
3155 | vector<uint8_t> os; | |
11fdf7f2 | 3156 | decode(os, bl); |
31f18b77 FG |
3157 | osd_state.resize(os.size()); |
3158 | for (unsigned i = 0; i < os.size(); ++i) { | |
3159 | osd_state[i] = os[i]; | |
3160 | } | |
3161 | } | |
11fdf7f2 TL |
3162 | decode(osd_weight, bl); |
3163 | decode(osd_addrs->client_addrs, bl); | |
7c673cae | 3164 | |
11fdf7f2 TL |
3165 | decode(*pg_temp, bl); |
3166 | decode(*primary_temp, bl); | |
3167 | // dates back to firefly. version increased from 2 to 3 still in firefly. | |
3168 | // do we really still need to keep this around? even for old clients? | |
7c673cae FG |
3169 | if (struct_v >= 2) { |
3170 | osd_primary_affinity.reset(new mempool::osdmap::vector<__u32>); | |
11fdf7f2 | 3171 | decode(*osd_primary_affinity, bl); |
7c673cae FG |
3172 | if (osd_primary_affinity->empty()) |
3173 | osd_primary_affinity.reset(); | |
3174 | } else { | |
3175 | osd_primary_affinity.reset(); | |
3176 | } | |
3177 | ||
3178 | // crush | |
3179 | bufferlist cbl; | |
11fdf7f2 TL |
3180 | decode(cbl, bl); |
3181 | auto cblp = cbl.cbegin(); | |
7c673cae | 3182 | crush->decode(cblp); |
11fdf7f2 TL |
3183 | // added in firefly; version increased in luminous, so it affects |
3184 | // giant, hammer, infernallis, jewel, and kraken. probably should be left | |
3185 | // alone until we require clients to be all luminous? | |
7c673cae | 3186 | if (struct_v >= 3) { |
11fdf7f2 | 3187 | decode(erasure_code_profiles, bl); |
7c673cae FG |
3188 | } else { |
3189 | erasure_code_profiles.clear(); | |
3190 | } | |
11fdf7f2 TL |
3191 | // version increased from 3 to 4 still in luminous, so same as above |
3192 | // applies. | |
7c673cae | 3193 | if (struct_v >= 4) { |
11fdf7f2 TL |
3194 | decode(pg_upmap, bl); |
3195 | decode(pg_upmap_items, bl); | |
7c673cae FG |
3196 | } else { |
3197 | pg_upmap.clear(); | |
3198 | pg_upmap_items.clear(); | |
3199 | } | |
11fdf7f2 TL |
3200 | // again, version increased from 5 to 6 still in luminous, so above |
3201 | // applies. | |
31f18b77 | 3202 | if (struct_v >= 6) { |
11fdf7f2 TL |
3203 | decode(crush_version, bl); |
3204 | } | |
3205 | // version increase from 6 to 7 in mimic | |
3206 | if (struct_v >= 7) { | |
3207 | decode(new_removed_snaps, bl); | |
3208 | decode(new_purged_snaps, bl); | |
3209 | } | |
3210 | // version increase from 7 to 8, 8 to 9, in nautilus. | |
3211 | if (struct_v >= 9) { | |
3212 | decode(last_up_change, bl); | |
3213 | decode(last_in_change, bl); | |
31f18b77 | 3214 | } |
7c673cae FG |
3215 | DECODE_FINISH(bl); // client-usable data |
3216 | } | |
3217 | ||
3218 | { | |
81eedcae | 3219 | DECODE_START(9, bl); // extended, osd-only data |
11fdf7f2 TL |
3220 | decode(osd_addrs->hb_back_addrs, bl); |
3221 | decode(osd_info, bl); | |
3222 | decode(blacklist, bl); | |
3223 | decode(osd_addrs->cluster_addrs, bl); | |
3224 | decode(cluster_snapshot_epoch, bl); | |
3225 | decode(cluster_snapshot, bl); | |
3226 | decode(*osd_uuid, bl); | |
3227 | decode(osd_xinfo, bl); | |
3228 | decode(osd_addrs->hb_front_addrs, bl); | |
3229 | // | |
7c673cae | 3230 | if (struct_v >= 2) { |
11fdf7f2 TL |
3231 | decode(nearfull_ratio, bl); |
3232 | decode(full_ratio, bl); | |
7c673cae FG |
3233 | } else { |
3234 | nearfull_ratio = 0; | |
3235 | full_ratio = 0; | |
3236 | } | |
3237 | if (struct_v >= 3) { | |
11fdf7f2 | 3238 | decode(backfillfull_ratio, bl); |
7c673cae FG |
3239 | } else { |
3240 | backfillfull_ratio = 0; | |
3241 | } | |
31f18b77 FG |
3242 | if (struct_v == 4) { |
3243 | string r; | |
11fdf7f2 | 3244 | decode(r, bl); |
31f18b77 FG |
3245 | if (r.length()) |
3246 | require_min_compat_client = ceph_release_from_name(r.c_str()); | |
3247 | } | |
3248 | if (struct_v >= 5) { | |
11fdf7f2 TL |
3249 | decode(require_min_compat_client, bl); |
3250 | decode(require_osd_release, bl); | |
3251 | if (require_osd_release >= CEPH_RELEASE_NAUTILUS) { | |
3252 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; | |
3253 | } | |
31f18b77 FG |
3254 | if (require_osd_release >= CEPH_RELEASE_LUMINOUS) { |
3255 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 3256 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
3257 | } |
3258 | } else { | |
3259 | if (flags & CEPH_OSDMAP_REQUIRE_LUMINOUS) { | |
3260 | // only for compat with post-kraken pre-luminous test clusters | |
3261 | require_osd_release = CEPH_RELEASE_LUMINOUS; | |
3262 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); | |
c07f9fc5 | 3263 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
3264 | } else if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { |
3265 | require_osd_release = CEPH_RELEASE_KRAKEN; | |
3266 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
3267 | require_osd_release = CEPH_RELEASE_JEWEL; | |
3268 | } else { | |
3269 | require_osd_release = 0; | |
3270 | } | |
3271 | } | |
11fdf7f2 TL |
3272 | if (struct_v >= 6) { |
3273 | decode(removed_snaps_queue, bl); | |
3274 | } | |
81eedcae TL |
3275 | if (struct_v >= 8) { |
3276 | decode(crush_node_flags, bl); | |
3277 | } else { | |
3278 | crush_node_flags.clear(); | |
3279 | } | |
3280 | if (struct_v >= 9) { | |
3281 | decode(device_class_flags, bl); | |
3282 | } else { | |
3283 | device_class_flags.clear(); | |
3284 | } | |
7c673cae FG |
3285 | DECODE_FINISH(bl); // osd-only data |
3286 | } | |
3287 | ||
3288 | if (struct_v >= 8) { | |
3289 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 3290 | decode(crc, bl); |
7c673cae FG |
3291 | tail_offset = bl.get_off(); |
3292 | crc_defined = true; | |
3293 | } else { | |
3294 | crc_defined = false; | |
3295 | crc = 0; | |
3296 | } | |
3297 | ||
3298 | DECODE_FINISH(bl); // wrapper | |
3299 | ||
3300 | if (tail_offset) { | |
3301 | // verify crc | |
3302 | uint32_t actual = crc_front.crc32c(-1); | |
3303 | if (tail_offset < bl.get_off()) { | |
3304 | bufferlist tail; | |
3305 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); | |
3306 | actual = tail.crc32c(actual); | |
3307 | } | |
3308 | if (crc != actual) { | |
3309 | ostringstream ss; | |
3310 | ss << "bad crc, actual " << actual << " != expected " << crc; | |
3311 | string s = ss.str(); | |
3312 | throw buffer::malformed_input(s.c_str()); | |
3313 | } | |
3314 | } | |
3315 | ||
3316 | post_decode(); | |
3317 | } | |
3318 | ||
3319 | void OSDMap::post_decode() | |
3320 | { | |
3321 | // index pool names | |
3322 | name_pool.clear(); | |
3323 | for (const auto &pname : pool_name) { | |
3324 | name_pool[pname.second] = pname.first; | |
3325 | } | |
3326 | ||
3327 | calc_num_osds(); | |
3328 | _calc_up_osd_features(); | |
3329 | } | |
3330 | ||
3331 | void OSDMap::dump_erasure_code_profiles( | |
3332 | const mempool::osdmap::map<string,map<string,string>>& profiles, | |
3333 | Formatter *f) | |
3334 | { | |
3335 | f->open_object_section("erasure_code_profiles"); | |
3336 | for (const auto &profile : profiles) { | |
3337 | f->open_object_section(profile.first.c_str()); | |
3338 | for (const auto &profm : profile.second) { | |
3339 | f->dump_string(profm.first.c_str(), profm.second.c_str()); | |
3340 | } | |
3341 | f->close_section(); | |
3342 | } | |
3343 | f->close_section(); | |
3344 | } | |
3345 | ||
3346 | void OSDMap::dump(Formatter *f) const | |
3347 | { | |
3348 | f->dump_int("epoch", get_epoch()); | |
3349 | f->dump_stream("fsid") << get_fsid(); | |
3350 | f->dump_stream("created") << get_created(); | |
3351 | f->dump_stream("modified") << get_modified(); | |
11fdf7f2 TL |
3352 | f->dump_stream("last_up_change") << last_up_change; |
3353 | f->dump_stream("last_in_change") << last_in_change; | |
7c673cae | 3354 | f->dump_string("flags", get_flag_string()); |
11fdf7f2 TL |
3355 | f->dump_unsigned("flags_num", flags); |
3356 | f->open_array_section("flags_set"); | |
3357 | set<string> flagset; | |
3358 | get_flag_set(&flagset); | |
3359 | for (auto p : flagset) { | |
3360 | f->dump_string("flag", p); | |
3361 | } | |
3362 | f->close_section(); | |
31f18b77 | 3363 | f->dump_unsigned("crush_version", get_crush_version()); |
7c673cae FG |
3364 | f->dump_float("full_ratio", full_ratio); |
3365 | f->dump_float("backfillfull_ratio", backfillfull_ratio); | |
3366 | f->dump_float("nearfull_ratio", nearfull_ratio); | |
3367 | f->dump_string("cluster_snapshot", get_cluster_snapshot()); | |
3368 | f->dump_int("pool_max", get_pool_max()); | |
3369 | f->dump_int("max_osd", get_max_osd()); | |
31f18b77 FG |
3370 | f->dump_string("require_min_compat_client", |
3371 | ceph_release_name(require_min_compat_client)); | |
3372 | f->dump_string("min_compat_client", | |
3373 | ceph_release_name(get_min_compat_client())); | |
3374 | f->dump_string("require_osd_release", | |
3375 | ceph_release_name(require_osd_release)); | |
7c673cae FG |
3376 | |
3377 | f->open_array_section("pools"); | |
3378 | for (const auto &pool : pools) { | |
3379 | std::string name("<unknown>"); | |
3380 | const auto &pni = pool_name.find(pool.first); | |
3381 | if (pni != pool_name.end()) | |
3382 | name = pni->second; | |
3383 | f->open_object_section("pool"); | |
3384 | f->dump_int("pool", pool.first); | |
3385 | f->dump_string("pool_name", name); | |
3386 | pool.second.dump(f); | |
3387 | f->close_section(); | |
3388 | } | |
3389 | f->close_section(); | |
3390 | ||
3391 | f->open_array_section("osds"); | |
3392 | for (int i=0; i<get_max_osd(); i++) | |
3393 | if (exists(i)) { | |
3394 | f->open_object_section("osd_info"); | |
3395 | f->dump_int("osd", i); | |
3396 | f->dump_stream("uuid") << get_uuid(i); | |
3397 | f->dump_int("up", is_up(i)); | |
3398 | f->dump_int("in", is_in(i)); | |
3399 | f->dump_float("weight", get_weightf(i)); | |
3400 | f->dump_float("primary_affinity", get_primary_affinityf(i)); | |
3401 | get_info(i).dump(f); | |
11fdf7f2 TL |
3402 | f->dump_object("public_addrs", get_addrs(i)); |
3403 | f->dump_object("cluster_addrs", get_cluster_addrs(i)); | |
3404 | f->dump_object("heartbeat_back_addrs", get_hb_back_addrs(i)); | |
3405 | f->dump_object("heartbeat_front_addrs", get_hb_front_addrs(i)); | |
3406 | // compat | |
3407 | f->dump_stream("public_addr") << get_addrs(i).get_legacy_str(); | |
3408 | f->dump_stream("cluster_addr") << get_cluster_addrs(i).get_legacy_str(); | |
3409 | f->dump_stream("heartbeat_back_addr") | |
3410 | << get_hb_back_addrs(i).get_legacy_str(); | |
3411 | f->dump_stream("heartbeat_front_addr") | |
3412 | << get_hb_front_addrs(i).get_legacy_str(); | |
7c673cae FG |
3413 | |
3414 | set<string> st; | |
3415 | get_state(i, st); | |
3416 | f->open_array_section("state"); | |
3417 | for (const auto &state : st) | |
3418 | f->dump_string("state", state); | |
3419 | f->close_section(); | |
3420 | ||
3421 | f->close_section(); | |
3422 | } | |
3423 | f->close_section(); | |
3424 | ||
3425 | f->open_array_section("osd_xinfo"); | |
3426 | for (int i=0; i<get_max_osd(); i++) { | |
3427 | if (exists(i)) { | |
3428 | f->open_object_section("xinfo"); | |
3429 | f->dump_int("osd", i); | |
3430 | osd_xinfo[i].dump(f); | |
3431 | f->close_section(); | |
3432 | } | |
3433 | } | |
3434 | f->close_section(); | |
3435 | ||
3436 | f->open_array_section("pg_upmap"); | |
3437 | for (auto& p : pg_upmap) { | |
3438 | f->open_object_section("mapping"); | |
3439 | f->dump_stream("pgid") << p.first; | |
3440 | f->open_array_section("osds"); | |
3441 | for (auto q : p.second) { | |
3442 | f->dump_int("osd", q); | |
3443 | } | |
3444 | f->close_section(); | |
3445 | f->close_section(); | |
3446 | } | |
3447 | f->close_section(); | |
3448 | f->open_array_section("pg_upmap_items"); | |
3449 | for (auto& p : pg_upmap_items) { | |
3450 | f->open_object_section("mapping"); | |
3451 | f->dump_stream("pgid") << p.first; | |
3452 | f->open_array_section("mappings"); | |
3453 | for (auto& q : p.second) { | |
3454 | f->open_object_section("mapping"); | |
3455 | f->dump_int("from", q.first); | |
3456 | f->dump_int("to", q.second); | |
3457 | f->close_section(); | |
3458 | } | |
3459 | f->close_section(); | |
3460 | f->close_section(); | |
3461 | } | |
3462 | f->close_section(); | |
3463 | f->open_array_section("pg_temp"); | |
31f18b77 | 3464 | pg_temp->dump(f); |
7c673cae FG |
3465 | f->close_section(); |
3466 | ||
3467 | f->open_array_section("primary_temp"); | |
3468 | for (const auto &pg : *primary_temp) { | |
3469 | f->dump_stream("pgid") << pg.first; | |
3470 | f->dump_int("osd", pg.second); | |
3471 | } | |
3472 | f->close_section(); // primary_temp | |
3473 | ||
3474 | f->open_object_section("blacklist"); | |
3475 | for (const auto &addr : blacklist) { | |
3476 | stringstream ss; | |
3477 | ss << addr.first; | |
3478 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3479 | } | |
3480 | f->close_section(); | |
3481 | ||
3482 | dump_erasure_code_profiles(erasure_code_profiles, f); | |
11fdf7f2 TL |
3483 | |
3484 | f->open_array_section("removed_snaps_queue"); | |
3485 | for (auto& p : removed_snaps_queue) { | |
3486 | f->open_object_section("pool"); | |
3487 | f->dump_int("pool", p.first); | |
3488 | f->open_array_section("snaps"); | |
3489 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3490 | f->open_object_section("interval"); | |
3491 | f->dump_unsigned("begin", q.get_start()); | |
3492 | f->dump_unsigned("length", q.get_len()); | |
3493 | f->close_section(); | |
3494 | } | |
3495 | f->close_section(); | |
3496 | f->close_section(); | |
3497 | } | |
3498 | f->close_section(); | |
3499 | f->open_array_section("new_removed_snaps"); | |
3500 | for (auto& p : new_removed_snaps) { | |
3501 | f->open_object_section("pool"); | |
3502 | f->dump_int("pool", p.first); | |
3503 | f->open_array_section("snaps"); | |
3504 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3505 | f->open_object_section("interval"); | |
3506 | f->dump_unsigned("begin", q.get_start()); | |
3507 | f->dump_unsigned("length", q.get_len()); | |
3508 | f->close_section(); | |
3509 | } | |
3510 | f->close_section(); | |
3511 | f->close_section(); | |
3512 | } | |
3513 | f->close_section(); | |
3514 | f->open_array_section("new_purged_snaps"); | |
3515 | for (auto& p : new_purged_snaps) { | |
3516 | f->open_object_section("pool"); | |
3517 | f->dump_int("pool", p.first); | |
3518 | f->open_array_section("snaps"); | |
3519 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3520 | f->open_object_section("interval"); | |
3521 | f->dump_unsigned("begin", q.get_start()); | |
3522 | f->dump_unsigned("length", q.get_len()); | |
3523 | f->close_section(); | |
3524 | } | |
3525 | f->close_section(); | |
3526 | f->close_section(); | |
3527 | } | |
3528 | f->close_section(); | |
81eedcae TL |
3529 | f->open_object_section("crush_node_flags"); |
3530 | for (auto& i : crush_node_flags) { | |
3531 | string s = crush->item_exists(i.first) ? crush->get_item_name(i.first) | |
3532 | : stringify(i.first); | |
3533 | f->open_array_section(s.c_str()); | |
3534 | set<string> st; | |
3535 | calc_state_set(i.second, st); | |
3536 | for (auto& j : st) { | |
3537 | f->dump_string("flag", j); | |
3538 | } | |
3539 | f->close_section(); | |
3540 | } | |
3541 | f->close_section(); | |
3542 | f->open_object_section("device_class_flags"); | |
3543 | for (auto& i : device_class_flags) { | |
3544 | const char* class_name = crush->get_class_name(i.first); | |
3545 | string s = class_name ? class_name : stringify(i.first); | |
3546 | f->open_array_section(s.c_str()); | |
3547 | set<string> st; | |
3548 | calc_state_set(i.second, st); | |
3549 | for (auto& j : st) { | |
3550 | f->dump_string("flag", j); | |
3551 | } | |
3552 | f->close_section(); | |
3553 | } | |
3554 | f->close_section(); | |
7c673cae FG |
3555 | } |
3556 | ||
3557 | void OSDMap::generate_test_instances(list<OSDMap*>& o) | |
3558 | { | |
3559 | o.push_back(new OSDMap); | |
3560 | ||
3561 | CephContext *cct = new CephContext(CODE_ENVIRONMENT_UTILITY); | |
3562 | o.push_back(new OSDMap); | |
3563 | uuid_d fsid; | |
224ce89b | 3564 | o.back()->build_simple(cct, 1, fsid, 16); |
7c673cae FG |
3565 | o.back()->created = o.back()->modified = utime_t(1, 2); // fix timestamp |
3566 | o.back()->blacklist[entity_addr_t()] = utime_t(5, 6); | |
3567 | cct->put(); | |
3568 | } | |
3569 | ||
3570 | string OSDMap::get_flag_string(unsigned f) | |
3571 | { | |
3572 | string s; | |
3573 | if ( f& CEPH_OSDMAP_NEARFULL) | |
3574 | s += ",nearfull"; | |
3575 | if (f & CEPH_OSDMAP_FULL) | |
3576 | s += ",full"; | |
3577 | if (f & CEPH_OSDMAP_PAUSERD) | |
3578 | s += ",pauserd"; | |
3579 | if (f & CEPH_OSDMAP_PAUSEWR) | |
3580 | s += ",pausewr"; | |
3581 | if (f & CEPH_OSDMAP_PAUSEREC) | |
3582 | s += ",pauserec"; | |
3583 | if (f & CEPH_OSDMAP_NOUP) | |
3584 | s += ",noup"; | |
3585 | if (f & CEPH_OSDMAP_NODOWN) | |
3586 | s += ",nodown"; | |
3587 | if (f & CEPH_OSDMAP_NOOUT) | |
3588 | s += ",noout"; | |
3589 | if (f & CEPH_OSDMAP_NOIN) | |
3590 | s += ",noin"; | |
3591 | if (f & CEPH_OSDMAP_NOBACKFILL) | |
3592 | s += ",nobackfill"; | |
3593 | if (f & CEPH_OSDMAP_NOREBALANCE) | |
3594 | s += ",norebalance"; | |
3595 | if (f & CEPH_OSDMAP_NORECOVER) | |
3596 | s += ",norecover"; | |
3597 | if (f & CEPH_OSDMAP_NOSCRUB) | |
3598 | s += ",noscrub"; | |
3599 | if (f & CEPH_OSDMAP_NODEEP_SCRUB) | |
3600 | s += ",nodeep-scrub"; | |
3601 | if (f & CEPH_OSDMAP_NOTIERAGENT) | |
3602 | s += ",notieragent"; | |
11fdf7f2 TL |
3603 | if (f & CEPH_OSDMAP_NOSNAPTRIM) |
3604 | s += ",nosnaptrim"; | |
7c673cae FG |
3605 | if (f & CEPH_OSDMAP_SORTBITWISE) |
3606 | s += ",sortbitwise"; | |
3607 | if (f & CEPH_OSDMAP_REQUIRE_JEWEL) | |
3608 | s += ",require_jewel_osds"; | |
3609 | if (f & CEPH_OSDMAP_REQUIRE_KRAKEN) | |
3610 | s += ",require_kraken_osds"; | |
3611 | if (f & CEPH_OSDMAP_REQUIRE_LUMINOUS) | |
3612 | s += ",require_luminous_osds"; | |
c07f9fc5 FG |
3613 | if (f & CEPH_OSDMAP_RECOVERY_DELETES) |
3614 | s += ",recovery_deletes"; | |
181888fb FG |
3615 | if (f & CEPH_OSDMAP_PURGED_SNAPDIRS) |
3616 | s += ",purged_snapdirs"; | |
f64942e4 AA |
3617 | if (f & CEPH_OSDMAP_PGLOG_HARDLIMIT) |
3618 | s += ",pglog_hardlimit"; | |
7c673cae FG |
3619 | if (s.length()) |
3620 | s.erase(0, 1); | |
3621 | return s; | |
3622 | } | |
3623 | ||
3624 | string OSDMap::get_flag_string() const | |
3625 | { | |
3626 | return get_flag_string(flags); | |
3627 | } | |
3628 | ||
7c673cae FG |
3629 | void OSDMap::print_pools(ostream& out) const |
3630 | { | |
3631 | for (const auto &pool : pools) { | |
3632 | std::string name("<unknown>"); | |
3633 | const auto &pni = pool_name.find(pool.first); | |
3634 | if (pni != pool_name.end()) | |
3635 | name = pni->second; | |
3636 | out << "pool " << pool.first | |
3637 | << " '" << name | |
3638 | << "' " << pool.second << "\n"; | |
3639 | ||
3640 | for (const auto &snap : pool.second.snaps) | |
3641 | out << "\tsnap " << snap.second.snapid << " '" << snap.second.name << "' " << snap.second.stamp << "\n"; | |
3642 | ||
3643 | if (!pool.second.removed_snaps.empty()) | |
3644 | out << "\tremoved_snaps " << pool.second.removed_snaps << "\n"; | |
11fdf7f2 TL |
3645 | auto p = removed_snaps_queue.find(pool.first); |
3646 | if (p != removed_snaps_queue.end()) { | |
3647 | out << "\tremoved_snaps_queue " << p->second << "\n"; | |
3648 | } | |
7c673cae FG |
3649 | } |
3650 | out << std::endl; | |
3651 | } | |
3652 | ||
3653 | void OSDMap::print(ostream& out) const | |
3654 | { | |
3655 | out << "epoch " << get_epoch() << "\n" | |
3656 | << "fsid " << get_fsid() << "\n" | |
3657 | << "created " << get_created() << "\n" | |
3658 | << "modified " << get_modified() << "\n"; | |
3659 | ||
3660 | out << "flags " << get_flag_string() << "\n"; | |
31f18b77 | 3661 | out << "crush_version " << get_crush_version() << "\n"; |
7c673cae FG |
3662 | out << "full_ratio " << full_ratio << "\n"; |
3663 | out << "backfillfull_ratio " << backfillfull_ratio << "\n"; | |
3664 | out << "nearfull_ratio " << nearfull_ratio << "\n"; | |
31f18b77 FG |
3665 | if (require_min_compat_client > 0) { |
3666 | out << "require_min_compat_client " | |
3667 | << ceph_release_name(require_min_compat_client) << "\n"; | |
7c673cae | 3668 | } |
31f18b77 FG |
3669 | out << "min_compat_client " << ceph_release_name(get_min_compat_client()) |
3670 | << "\n"; | |
224ce89b WB |
3671 | if (require_osd_release > 0) { |
3672 | out << "require_osd_release " << ceph_release_name(require_osd_release) | |
3673 | << "\n"; | |
3674 | } | |
7c673cae FG |
3675 | if (get_cluster_snapshot().length()) |
3676 | out << "cluster_snapshot " << get_cluster_snapshot() << "\n"; | |
3677 | out << "\n"; | |
3678 | ||
3679 | print_pools(out); | |
3680 | ||
3681 | out << "max_osd " << get_max_osd() << "\n"; | |
3682 | for (int i=0; i<get_max_osd(); i++) { | |
3683 | if (exists(i)) { | |
3684 | out << "osd." << i; | |
3685 | out << (is_up(i) ? " up ":" down"); | |
3686 | out << (is_in(i) ? " in ":" out"); | |
3687 | out << " weight " << get_weightf(i); | |
3688 | if (get_primary_affinity(i) != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) | |
3689 | out << " primary_affinity " << get_primary_affinityf(i); | |
3690 | const osd_info_t& info(get_info(i)); | |
3691 | out << " " << info; | |
11fdf7f2 | 3692 | out << " " << get_addrs(i) << " " << get_cluster_addrs(i); |
7c673cae FG |
3693 | set<string> st; |
3694 | get_state(i, st); | |
3695 | out << " " << st; | |
3696 | if (!get_uuid(i).is_zero()) | |
3697 | out << " " << get_uuid(i); | |
3698 | out << "\n"; | |
3699 | } | |
3700 | } | |
3701 | out << std::endl; | |
3702 | ||
3703 | for (auto& p : pg_upmap) { | |
3704 | out << "pg_upmap " << p.first << " " << p.second << "\n"; | |
3705 | } | |
3706 | for (auto& p : pg_upmap_items) { | |
3707 | out << "pg_upmap_items " << p.first << " " << p.second << "\n"; | |
3708 | } | |
3709 | ||
3710 | for (const auto pg : *pg_temp) | |
3711 | out << "pg_temp " << pg.first << " " << pg.second << "\n"; | |
3712 | ||
3713 | for (const auto pg : *primary_temp) | |
3714 | out << "primary_temp " << pg.first << " " << pg.second << "\n"; | |
3715 | ||
3716 | for (const auto &addr : blacklist) | |
3717 | out << "blacklist " << addr.first << " expires " << addr.second << "\n"; | |
7c673cae FG |
3718 | } |
3719 | ||
3720 | class OSDTreePlainDumper : public CrushTreeDumper::Dumper<TextTable> { | |
3721 | public: | |
3722 | typedef CrushTreeDumper::Dumper<TextTable> Parent; | |
31f18b77 FG |
3723 | |
3724 | OSDTreePlainDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
3725 | unsigned f) | |
c07f9fc5 | 3726 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
3727 | |
3728 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
3729 | if (!filter) { |
3730 | return true; // normal case | |
3731 | } | |
3732 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
3733 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
3734 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
3735 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
3736 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
3737 | return true; | |
31f18b77 | 3738 | } |
c07f9fc5 | 3739 | return false; |
31f18b77 FG |
3740 | } |
3741 | ||
3742 | bool should_dump_empty_bucket() const override { | |
3743 | return !filter; | |
3744 | } | |
7c673cae | 3745 | |
11fdf7f2 | 3746 | void init_table(TextTable *tbl) { |
7c673cae | 3747 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 3748 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
3749 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
3750 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
c07f9fc5 | 3751 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae | 3752 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 3753 | tbl->define_column("PRI-AFF", TextTable::LEFT, TextTable::RIGHT); |
11fdf7f2 TL |
3754 | } |
3755 | void dump(TextTable *tbl, string& bucket) { | |
3756 | init_table(tbl); | |
7c673cae | 3757 | |
11fdf7f2 TL |
3758 | if (!bucket.empty()) { |
3759 | set_root(bucket); | |
3760 | Parent::dump(tbl); | |
3761 | } else { | |
3762 | Parent::dump(tbl); | |
3763 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
3764 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) { | |
3765 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), tbl); | |
3766 | } | |
31f18b77 | 3767 | } |
7c673cae FG |
3768 | } |
3769 | } | |
3770 | ||
3771 | protected: | |
3772 | void dump_item(const CrushTreeDumper::Item &qi, TextTable *tbl) override { | |
224ce89b WB |
3773 | const char *c = crush->get_item_class(qi.id); |
3774 | if (!c) | |
3775 | c = ""; | |
7c673cae | 3776 | *tbl << qi.id |
224ce89b | 3777 | << c |
7c673cae FG |
3778 | << weightf_t(qi.weight); |
3779 | ||
3780 | ostringstream name; | |
3781 | for (int k = 0; k < qi.depth; k++) | |
3782 | name << " "; | |
3783 | if (qi.is_bucket()) { | |
3784 | name << crush->get_type_name(crush->get_bucket_type(qi.id)) << " " | |
3785 | << crush->get_item_name(qi.id); | |
3786 | } else { | |
3787 | name << "osd." << qi.id; | |
3788 | } | |
3789 | *tbl << name.str(); | |
3790 | ||
3791 | if (!qi.is_bucket()) { | |
3792 | if (!osdmap->exists(qi.id)) { | |
3793 | *tbl << "DNE" | |
3794 | << 0; | |
3795 | } else { | |
c07f9fc5 FG |
3796 | string s; |
3797 | if (osdmap->is_up(qi.id)) { | |
3798 | s = "up"; | |
3799 | } else if (osdmap->is_destroyed(qi.id)) { | |
3800 | s = "destroyed"; | |
3801 | } else { | |
3802 | s = "down"; | |
3803 | } | |
3804 | *tbl << s | |
7c673cae FG |
3805 | << weightf_t(osdmap->get_weightf(qi.id)) |
3806 | << weightf_t(osdmap->get_primary_affinityf(qi.id)); | |
3807 | } | |
3808 | } | |
3809 | *tbl << TextTable::endrow; | |
3810 | } | |
3811 | ||
3812 | private: | |
3813 | const OSDMap *osdmap; | |
31f18b77 | 3814 | const unsigned filter; |
7c673cae FG |
3815 | }; |
3816 | ||
3817 | class OSDTreeFormattingDumper : public CrushTreeDumper::FormattingDumper { | |
3818 | public: | |
3819 | typedef CrushTreeDumper::FormattingDumper Parent; | |
3820 | ||
31f18b77 FG |
3821 | OSDTreeFormattingDumper(const CrushWrapper *crush, const OSDMap *osdmap_, |
3822 | unsigned f) | |
c07f9fc5 | 3823 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
3824 | |
3825 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
3826 | if (!filter) { |
3827 | return true; // normal case | |
3828 | } | |
3829 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
3830 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
3831 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
3832 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
3833 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
3834 | return true; | |
31f18b77 | 3835 | } |
c07f9fc5 | 3836 | return false; |
31f18b77 FG |
3837 | } |
3838 | ||
3839 | bool should_dump_empty_bucket() const override { | |
3840 | return !filter; | |
3841 | } | |
7c673cae | 3842 | |
11fdf7f2 TL |
3843 | void dump(Formatter *f, string& bucket) { |
3844 | if (!bucket.empty()) { | |
3845 | set_root(bucket); | |
3846 | f->open_array_section("nodes"); | |
3847 | Parent::dump(f); | |
3848 | f->close_section(); | |
3849 | } else { | |
3850 | f->open_array_section("nodes"); | |
3851 | Parent::dump(f); | |
3852 | f->close_section(); | |
3853 | f->open_array_section("stray"); | |
3854 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
3855 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) | |
3856 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); | |
3857 | } | |
3858 | f->close_section(); | |
7c673cae | 3859 | } |
7c673cae FG |
3860 | } |
3861 | ||
3862 | protected: | |
3863 | void dump_item_fields(const CrushTreeDumper::Item &qi, Formatter *f) override { | |
3864 | Parent::dump_item_fields(qi, f); | |
3865 | if (!qi.is_bucket()) | |
3866 | { | |
c07f9fc5 FG |
3867 | string s; |
3868 | if (osdmap->is_up(qi.id)) { | |
3869 | s = "up"; | |
3870 | } else if (osdmap->is_destroyed(qi.id)) { | |
3871 | s = "destroyed"; | |
3872 | } else { | |
3873 | s = "down"; | |
3874 | } | |
7c673cae | 3875 | f->dump_unsigned("exists", (int)osdmap->exists(qi.id)); |
c07f9fc5 | 3876 | f->dump_string("status", s); |
7c673cae FG |
3877 | f->dump_float("reweight", osdmap->get_weightf(qi.id)); |
3878 | f->dump_float("primary_affinity", osdmap->get_primary_affinityf(qi.id)); | |
3879 | } | |
3880 | } | |
3881 | ||
3882 | private: | |
3883 | const OSDMap *osdmap; | |
31f18b77 | 3884 | const unsigned filter; |
7c673cae FG |
3885 | }; |
3886 | ||
11fdf7f2 | 3887 | void OSDMap::print_tree(Formatter *f, ostream *out, unsigned filter, string bucket) const |
7c673cae | 3888 | { |
31f18b77 | 3889 | if (f) { |
11fdf7f2 | 3890 | OSDTreeFormattingDumper(crush.get(), this, filter).dump(f, bucket); |
31f18b77 | 3891 | } else { |
11fdf7f2 | 3892 | ceph_assert(out); |
7c673cae | 3893 | TextTable tbl; |
11fdf7f2 | 3894 | OSDTreePlainDumper(crush.get(), this, filter).dump(&tbl, bucket); |
7c673cae FG |
3895 | *out << tbl; |
3896 | } | |
3897 | } | |
3898 | ||
224ce89b | 3899 | void OSDMap::print_summary(Formatter *f, ostream& out, |
11fdf7f2 | 3900 | const string& prefix, bool extra) const |
7c673cae FG |
3901 | { |
3902 | if (f) { | |
3903 | f->open_object_section("osdmap"); | |
3904 | f->dump_int("epoch", get_epoch()); | |
3905 | f->dump_int("num_osds", get_num_osds()); | |
3906 | f->dump_int("num_up_osds", get_num_up_osds()); | |
3907 | f->dump_int("num_in_osds", get_num_in_osds()); | |
7c673cae FG |
3908 | f->dump_unsigned("num_remapped_pgs", get_num_pg_temp()); |
3909 | f->close_section(); | |
3910 | } else { | |
11fdf7f2 | 3911 | utime_t now = ceph_clock_now(); |
31f18b77 | 3912 | out << get_num_osds() << " osds: " |
11fdf7f2 TL |
3913 | << get_num_up_osds() << " up"; |
3914 | if (last_up_change != utime_t()) { | |
3915 | out << " (since " << utimespan_str(now - last_up_change) << ")"; | |
3916 | } | |
3917 | out << ", " << get_num_in_osds() << " in"; | |
3918 | if (last_in_change != utime_t()) { | |
3919 | out << " (since " << utimespan_str(now - last_in_change) << ")"; | |
3920 | } | |
3921 | if (extra) | |
3922 | out << "; epoch: e" << get_epoch(); | |
7c673cae FG |
3923 | if (get_num_pg_temp()) |
3924 | out << "; " << get_num_pg_temp() << " remapped pgs"; | |
3925 | out << "\n"; | |
3926 | uint64_t important_flags = flags & ~CEPH_OSDMAP_SEMIHIDDEN_FLAGS; | |
3927 | if (important_flags) | |
224ce89b | 3928 | out << prefix << "flags " << get_flag_string(important_flags) << "\n"; |
7c673cae FG |
3929 | } |
3930 | } | |
3931 | ||
3932 | void OSDMap::print_oneline_summary(ostream& out) const | |
3933 | { | |
3934 | out << "e" << get_epoch() << ": " | |
31f18b77 | 3935 | << get_num_osds() << " total, " |
7c673cae FG |
3936 | << get_num_up_osds() << " up, " |
3937 | << get_num_in_osds() << " in"; | |
7c673cae FG |
3938 | } |
3939 | ||
3efd9988 | 3940 | bool OSDMap::crush_rule_in_use(int rule_id) const |
7c673cae FG |
3941 | { |
3942 | for (const auto &pool : pools) { | |
3efd9988 | 3943 | if (pool.second.crush_rule == rule_id) |
7c673cae FG |
3944 | return true; |
3945 | } | |
3946 | return false; | |
3947 | } | |
3948 | ||
3efd9988 FG |
3949 | int OSDMap::validate_crush_rules(CrushWrapper *newcrush, |
3950 | ostream *ss) const | |
3951 | { | |
3952 | for (auto& i : pools) { | |
3953 | auto& pool = i.second; | |
3954 | int ruleno = pool.get_crush_rule(); | |
3955 | if (!newcrush->rule_exists(ruleno)) { | |
3956 | *ss << "pool " << i.first << " references crush_rule " << ruleno | |
3957 | << " but it is not present"; | |
3958 | return -EINVAL; | |
3959 | } | |
3960 | if (newcrush->get_rule_mask_ruleset(ruleno) != ruleno) { | |
3961 | *ss << "rule " << ruleno << " mask ruleset does not match rule id"; | |
3962 | return -EINVAL; | |
3963 | } | |
3964 | if (newcrush->get_rule_mask_type(ruleno) != (int)pool.get_type()) { | |
3965 | *ss << "pool " << i.first << " type does not match rule " << ruleno; | |
3966 | return -EINVAL; | |
3967 | } | |
11fdf7f2 TL |
3968 | int poolsize = pool.get_size(); |
3969 | if (poolsize < newcrush->get_rule_mask_min_size(ruleno) || | |
3970 | poolsize > newcrush->get_rule_mask_max_size(ruleno)) { | |
3971 | *ss << "pool " << i.first << " size " << poolsize << " does not" | |
3efd9988 FG |
3972 | << " fall within rule " << ruleno |
3973 | << " min_size " << newcrush->get_rule_mask_min_size(ruleno) | |
3974 | << " and max_size " << newcrush->get_rule_mask_max_size(ruleno); | |
3975 | return -EINVAL; | |
3976 | } | |
3977 | } | |
3978 | return 0; | |
3979 | } | |
3980 | ||
224ce89b WB |
3981 | int OSDMap::build_simple_optioned(CephContext *cct, epoch_t e, uuid_d &fsid, |
3982 | int nosd, int pg_bits, int pgp_bits, | |
3983 | bool default_pool) | |
7c673cae | 3984 | { |
224ce89b WB |
3985 | ldout(cct, 10) << "build_simple on " << nosd |
3986 | << " osds" << dendl; | |
7c673cae FG |
3987 | epoch = e; |
3988 | set_fsid(fsid); | |
3989 | created = modified = ceph_clock_now(); | |
3990 | ||
3991 | if (nosd >= 0) { | |
3992 | set_max_osd(nosd); | |
3993 | } else { | |
3994 | // count osds | |
3995 | int maxosd = 0; | |
11fdf7f2 | 3996 | const auto& conf = cct->_conf; |
7c673cae | 3997 | vector<string> sections; |
11fdf7f2 | 3998 | conf.get_all_sections(sections); |
7c673cae FG |
3999 | |
4000 | for (auto §ion : sections) { | |
4001 | if (section.find("osd.") != 0) | |
4002 | continue; | |
4003 | ||
4004 | const char *begin = section.c_str() + 4; | |
4005 | char *end = (char*)begin; | |
4006 | int o = strtol(begin, &end, 10); | |
4007 | if (*end != '\0') | |
4008 | continue; | |
4009 | ||
4010 | if (o > cct->_conf->mon_max_osd) { | |
4011 | lderr(cct) << "[osd." << o << "] in config has id > mon_max_osd " << cct->_conf->mon_max_osd << dendl; | |
4012 | return -ERANGE; | |
4013 | } | |
4014 | ||
4015 | if (o > maxosd) | |
4016 | maxosd = o; | |
4017 | } | |
4018 | ||
4019 | set_max_osd(maxosd + 1); | |
4020 | } | |
4021 | ||
7c673cae FG |
4022 | |
4023 | stringstream ss; | |
4024 | int r; | |
4025 | if (nosd >= 0) | |
4026 | r = build_simple_crush_map(cct, *crush, nosd, &ss); | |
4027 | else | |
4028 | r = build_simple_crush_map_from_conf(cct, *crush, &ss); | |
11fdf7f2 | 4029 | ceph_assert(r == 0); |
7c673cae FG |
4030 | |
4031 | int poolbase = get_max_osd() ? get_max_osd() : 1; | |
4032 | ||
d2e6a577 | 4033 | const int default_replicated_rule = crush->get_osd_pool_default_crush_replicated_ruleset(cct); |
11fdf7f2 | 4034 | ceph_assert(default_replicated_rule >= 0); |
7c673cae | 4035 | |
224ce89b WB |
4036 | if (default_pool) { |
4037 | // pgp_num <= pg_num | |
4038 | if (pgp_bits > pg_bits) | |
4039 | pgp_bits = pg_bits; | |
4040 | ||
4041 | vector<string> pool_names; | |
4042 | pool_names.push_back("rbd"); | |
4043 | for (auto &plname : pool_names) { | |
4044 | int64_t pool = ++pool_max; | |
4045 | pools[pool].type = pg_pool_t::TYPE_REPLICATED; | |
4046 | pools[pool].flags = cct->_conf->osd_pool_default_flags; | |
4047 | if (cct->_conf->osd_pool_default_flag_hashpspool) | |
4048 | pools[pool].set_flag(pg_pool_t::FLAG_HASHPSPOOL); | |
4049 | if (cct->_conf->osd_pool_default_flag_nodelete) | |
4050 | pools[pool].set_flag(pg_pool_t::FLAG_NODELETE); | |
4051 | if (cct->_conf->osd_pool_default_flag_nopgchange) | |
4052 | pools[pool].set_flag(pg_pool_t::FLAG_NOPGCHANGE); | |
4053 | if (cct->_conf->osd_pool_default_flag_nosizechange) | |
4054 | pools[pool].set_flag(pg_pool_t::FLAG_NOSIZECHANGE); | |
11fdf7f2 TL |
4055 | pools[pool].size = cct->_conf.get_val<uint64_t>("osd_pool_default_size"); |
4056 | pools[pool].min_size = cct->_conf.get_osd_pool_default_min_size( | |
4057 | pools[pool].size); | |
224ce89b WB |
4058 | pools[pool].crush_rule = default_replicated_rule; |
4059 | pools[pool].object_hash = CEPH_STR_HASH_RJENKINS; | |
4060 | pools[pool].set_pg_num(poolbase << pg_bits); | |
4061 | pools[pool].set_pgp_num(poolbase << pgp_bits); | |
11fdf7f2 TL |
4062 | pools[pool].set_pg_num_target(poolbase << pg_bits); |
4063 | pools[pool].set_pgp_num_target(poolbase << pgp_bits); | |
224ce89b | 4064 | pools[pool].last_change = epoch; |
c07f9fc5 FG |
4065 | pools[pool].application_metadata.insert( |
4066 | {pg_pool_t::APPLICATION_NAME_RBD, {}}); | |
11fdf7f2 TL |
4067 | auto m = pg_pool_t::get_pg_autoscale_mode_by_name( |
4068 | cct->_conf.get_val<string>("osd_pool_default_pg_autoscale_mode")); | |
4069 | pools[pool].pg_autoscale_mode = m >= 0 ? m : 0; | |
224ce89b WB |
4070 | pool_name[pool] = plname; |
4071 | name_pool[plname] = pool; | |
4072 | } | |
7c673cae FG |
4073 | } |
4074 | ||
4075 | for (int i=0; i<get_max_osd(); i++) { | |
4076 | set_state(i, 0); | |
4077 | set_weight(i, CEPH_OSD_OUT); | |
4078 | } | |
4079 | ||
4080 | map<string,string> profile_map; | |
4081 | r = get_erasure_code_profile_default(cct, profile_map, &ss); | |
4082 | if (r < 0) { | |
4083 | lderr(cct) << ss.str() << dendl; | |
4084 | return r; | |
4085 | } | |
4086 | set_erasure_code_profile("default", profile_map); | |
4087 | return 0; | |
4088 | } | |
4089 | ||
4090 | int OSDMap::get_erasure_code_profile_default(CephContext *cct, | |
4091 | map<string,string> &profile_map, | |
4092 | ostream *ss) | |
4093 | { | |
11fdf7f2 | 4094 | int r = get_json_str_map(cct->_conf.get_val<string>("osd_pool_default_erasure_code_profile"), |
7c673cae FG |
4095 | *ss, |
4096 | &profile_map); | |
4097 | return r; | |
4098 | } | |
4099 | ||
4100 | int OSDMap::_build_crush_types(CrushWrapper& crush) | |
4101 | { | |
4102 | crush.set_type_name(0, "osd"); | |
4103 | crush.set_type_name(1, "host"); | |
4104 | crush.set_type_name(2, "chassis"); | |
4105 | crush.set_type_name(3, "rack"); | |
4106 | crush.set_type_name(4, "row"); | |
4107 | crush.set_type_name(5, "pdu"); | |
4108 | crush.set_type_name(6, "pod"); | |
4109 | crush.set_type_name(7, "room"); | |
4110 | crush.set_type_name(8, "datacenter"); | |
11fdf7f2 TL |
4111 | crush.set_type_name(9, "zone"); |
4112 | crush.set_type_name(10, "region"); | |
4113 | crush.set_type_name(11, "root"); | |
4114 | return 11; | |
7c673cae FG |
4115 | } |
4116 | ||
4117 | int OSDMap::build_simple_crush_map(CephContext *cct, CrushWrapper& crush, | |
4118 | int nosd, ostream *ss) | |
4119 | { | |
4120 | crush.create(); | |
4121 | ||
4122 | // root | |
4123 | int root_type = _build_crush_types(crush); | |
4124 | int rootid; | |
4125 | int r = crush.add_bucket(0, 0, CRUSH_HASH_DEFAULT, | |
4126 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4127 | ceph_assert(r == 0); |
7c673cae FG |
4128 | crush.set_item_name(rootid, "default"); |
4129 | ||
4130 | for (int o=0; o<nosd; o++) { | |
4131 | map<string,string> loc; | |
4132 | loc["host"] = "localhost"; | |
4133 | loc["rack"] = "localrack"; | |
4134 | loc["root"] = "default"; | |
4135 | ldout(cct, 10) << " adding osd." << o << " at " << loc << dendl; | |
4136 | char name[32]; | |
4137 | snprintf(name, sizeof(name), "osd.%d", o); | |
4138 | crush.insert_item(cct, o, 1.0, name, loc); | |
4139 | } | |
4140 | ||
31f18b77 | 4141 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4142 | |
4143 | crush.finalize(); | |
4144 | ||
4145 | return 0; | |
4146 | } | |
4147 | ||
4148 | int OSDMap::build_simple_crush_map_from_conf(CephContext *cct, | |
4149 | CrushWrapper& crush, | |
4150 | ostream *ss) | |
4151 | { | |
11fdf7f2 | 4152 | const auto& conf = cct->_conf; |
7c673cae FG |
4153 | |
4154 | crush.create(); | |
4155 | ||
4156 | // root | |
4157 | int root_type = _build_crush_types(crush); | |
4158 | int rootid; | |
4159 | int r = crush.add_bucket(0, 0, | |
4160 | CRUSH_HASH_DEFAULT, | |
4161 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4162 | ceph_assert(r == 0); |
7c673cae FG |
4163 | crush.set_item_name(rootid, "default"); |
4164 | ||
4165 | // add osds | |
4166 | vector<string> sections; | |
11fdf7f2 | 4167 | conf.get_all_sections(sections); |
7c673cae FG |
4168 | |
4169 | for (auto §ion : sections) { | |
4170 | if (section.find("osd.") != 0) | |
4171 | continue; | |
4172 | ||
4173 | const char *begin = section.c_str() + 4; | |
4174 | char *end = (char*)begin; | |
4175 | int o = strtol(begin, &end, 10); | |
4176 | if (*end != '\0') | |
4177 | continue; | |
4178 | ||
4179 | string host, rack, row, room, dc, pool; | |
4180 | vector<string> sectiontmp; | |
4181 | sectiontmp.push_back("osd"); | |
4182 | sectiontmp.push_back(section); | |
11fdf7f2 TL |
4183 | conf.get_val_from_conf_file(sectiontmp, "host", host, false); |
4184 | conf.get_val_from_conf_file(sectiontmp, "rack", rack, false); | |
4185 | conf.get_val_from_conf_file(sectiontmp, "row", row, false); | |
4186 | conf.get_val_from_conf_file(sectiontmp, "room", room, false); | |
4187 | conf.get_val_from_conf_file(sectiontmp, "datacenter", dc, false); | |
4188 | conf.get_val_from_conf_file(sectiontmp, "root", pool, false); | |
7c673cae FG |
4189 | |
4190 | if (host.length() == 0) | |
4191 | host = "unknownhost"; | |
4192 | if (rack.length() == 0) | |
4193 | rack = "unknownrack"; | |
4194 | ||
4195 | map<string,string> loc; | |
4196 | loc["host"] = host; | |
4197 | loc["rack"] = rack; | |
4198 | if (row.size()) | |
4199 | loc["row"] = row; | |
4200 | if (room.size()) | |
4201 | loc["room"] = room; | |
4202 | if (dc.size()) | |
4203 | loc["datacenter"] = dc; | |
4204 | loc["root"] = "default"; | |
4205 | ||
4206 | ldout(cct, 5) << " adding osd." << o << " at " << loc << dendl; | |
4207 | crush.insert_item(cct, o, 1.0, section, loc); | |
4208 | } | |
4209 | ||
31f18b77 | 4210 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4211 | |
4212 | crush.finalize(); | |
4213 | ||
4214 | return 0; | |
4215 | } | |
4216 | ||
4217 | ||
31f18b77 FG |
4218 | int OSDMap::build_simple_crush_rules( |
4219 | CephContext *cct, | |
4220 | CrushWrapper& crush, | |
4221 | const string& root, | |
4222 | ostream *ss) | |
7c673cae | 4223 | { |
31f18b77 | 4224 | int crush_rule = crush.get_osd_pool_default_crush_replicated_ruleset(cct); |
7c673cae FG |
4225 | string failure_domain = |
4226 | crush.get_type_name(cct->_conf->osd_crush_chooseleaf_type); | |
4227 | ||
7c673cae | 4228 | int r; |
31f18b77 | 4229 | r = crush.add_simple_rule_at( |
224ce89b | 4230 | "replicated_rule", root, failure_domain, "", |
31f18b77 FG |
4231 | "firstn", pg_pool_t::TYPE_REPLICATED, |
4232 | crush_rule, ss); | |
7c673cae FG |
4233 | if (r < 0) |
4234 | return r; | |
4235 | // do not add an erasure rule by default or else we will implicitly | |
4236 | // require the crush_v2 feature of clients | |
4237 | return 0; | |
4238 | } | |
4239 | ||
4240 | int OSDMap::summarize_mapping_stats( | |
4241 | OSDMap *newmap, | |
4242 | const set<int64_t> *pools, | |
4243 | std::string *out, | |
4244 | Formatter *f) const | |
4245 | { | |
4246 | set<int64_t> ls; | |
4247 | if (pools) { | |
4248 | ls = *pools; | |
4249 | } else { | |
4250 | for (auto &p : get_pools()) | |
4251 | ls.insert(p.first); | |
4252 | } | |
4253 | ||
4254 | unsigned total_pg = 0; | |
4255 | unsigned moved_pg = 0; | |
4256 | vector<unsigned> base_by_osd(get_max_osd(), 0); | |
4257 | vector<unsigned> new_by_osd(get_max_osd(), 0); | |
4258 | for (int64_t pool_id : ls) { | |
4259 | const pg_pool_t *pi = get_pg_pool(pool_id); | |
31f18b77 FG |
4260 | vector<int> up, up2; |
4261 | int up_primary; | |
7c673cae | 4262 | for (unsigned ps = 0; ps < pi->get_pg_num(); ++ps) { |
11fdf7f2 | 4263 | pg_t pgid(ps, pool_id); |
7c673cae | 4264 | total_pg += pi->get_size(); |
31f18b77 | 4265 | pg_to_up_acting_osds(pgid, &up, &up_primary, nullptr, nullptr); |
7c673cae FG |
4266 | for (int osd : up) { |
4267 | if (osd >= 0 && osd < get_max_osd()) | |
4268 | ++base_by_osd[osd]; | |
4269 | } | |
4270 | if (newmap) { | |
31f18b77 | 4271 | newmap->pg_to_up_acting_osds(pgid, &up2, &up_primary, nullptr, nullptr); |
7c673cae FG |
4272 | for (int osd : up2) { |
4273 | if (osd >= 0 && osd < get_max_osd()) | |
4274 | ++new_by_osd[osd]; | |
4275 | } | |
4276 | if (pi->type == pg_pool_t::TYPE_ERASURE) { | |
4277 | for (unsigned i=0; i<up.size(); ++i) { | |
4278 | if (up[i] != up2[i]) { | |
4279 | ++moved_pg; | |
4280 | } | |
4281 | } | |
4282 | } else if (pi->type == pg_pool_t::TYPE_REPLICATED) { | |
4283 | for (int osd : up) { | |
4284 | if (std::find(up2.begin(), up2.end(), osd) == up2.end()) { | |
4285 | ++moved_pg; | |
4286 | } | |
4287 | } | |
4288 | } else { | |
11fdf7f2 | 4289 | ceph_abort_msg("unhandled pool type"); |
7c673cae FG |
4290 | } |
4291 | } | |
4292 | } | |
4293 | } | |
4294 | ||
4295 | unsigned num_up_in = 0; | |
4296 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4297 | if (is_up(osd) && is_in(osd)) | |
4298 | ++num_up_in; | |
4299 | } | |
4300 | if (!num_up_in) { | |
4301 | return -EINVAL; | |
4302 | } | |
4303 | ||
4304 | float avg_pg = (float)total_pg / (float)num_up_in; | |
4305 | float base_stddev = 0, new_stddev = 0; | |
4306 | int min = -1, max = -1; | |
4307 | unsigned min_base_pg = 0, max_base_pg = 0; | |
4308 | unsigned min_new_pg = 0, max_new_pg = 0; | |
4309 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4310 | if (is_up(osd) && is_in(osd)) { | |
4311 | float base_diff = (float)base_by_osd[osd] - avg_pg; | |
4312 | base_stddev += base_diff * base_diff; | |
4313 | float new_diff = (float)new_by_osd[osd] - avg_pg; | |
4314 | new_stddev += new_diff * new_diff; | |
4315 | if (min < 0 || base_by_osd[osd] < min_base_pg) { | |
4316 | min = osd; | |
4317 | min_base_pg = base_by_osd[osd]; | |
4318 | min_new_pg = new_by_osd[osd]; | |
4319 | } | |
4320 | if (max < 0 || base_by_osd[osd] > max_base_pg) { | |
4321 | max = osd; | |
4322 | max_base_pg = base_by_osd[osd]; | |
4323 | max_new_pg = new_by_osd[osd]; | |
4324 | } | |
4325 | } | |
4326 | } | |
4327 | base_stddev = sqrt(base_stddev / num_up_in); | |
4328 | new_stddev = sqrt(new_stddev / num_up_in); | |
4329 | ||
4330 | float edev = sqrt(avg_pg * (1.0 - (1.0 / (double)num_up_in))); | |
4331 | ||
4332 | ostringstream ss; | |
4333 | if (f) | |
4334 | f->open_object_section("utilization"); | |
4335 | if (newmap) { | |
4336 | if (f) { | |
4337 | f->dump_unsigned("moved_pgs", moved_pg); | |
4338 | f->dump_unsigned("total_pgs", total_pg); | |
4339 | } else { | |
4340 | float percent = 0; | |
4341 | if (total_pg) | |
4342 | percent = (float)moved_pg * 100.0 / (float)total_pg; | |
4343 | ss << "moved " << moved_pg << " / " << total_pg | |
4344 | << " (" << percent << "%)\n"; | |
4345 | } | |
4346 | } | |
4347 | if (f) { | |
4348 | f->dump_float("avg_pgs", avg_pg); | |
4349 | f->dump_float("std_dev", base_stddev); | |
4350 | f->dump_float("expected_baseline_std_dev", edev); | |
4351 | if (newmap) | |
4352 | f->dump_float("new_std_dev", new_stddev); | |
4353 | } else { | |
4354 | ss << "avg " << avg_pg << "\n"; | |
4355 | ss << "stddev " << base_stddev; | |
4356 | if (newmap) | |
4357 | ss << " -> " << new_stddev; | |
4358 | ss << " (expected baseline " << edev << ")\n"; | |
4359 | } | |
4360 | if (min >= 0) { | |
4361 | if (f) { | |
4362 | f->dump_unsigned("min_osd", min); | |
4363 | f->dump_unsigned("min_osd_pgs", min_base_pg); | |
4364 | if (newmap) | |
4365 | f->dump_unsigned("new_min_osd_pgs", min_new_pg); | |
4366 | } else { | |
4367 | ss << "min osd." << min << " with " << min_base_pg; | |
4368 | if (newmap) | |
4369 | ss << " -> " << min_new_pg; | |
4370 | ss << " pgs (" << (float)min_base_pg / avg_pg; | |
4371 | if (newmap) | |
4372 | ss << " -> " << (float)min_new_pg / avg_pg; | |
4373 | ss << " * mean)\n"; | |
4374 | } | |
4375 | } | |
4376 | if (max >= 0) { | |
4377 | if (f) { | |
4378 | f->dump_unsigned("max_osd", max); | |
4379 | f->dump_unsigned("max_osd_pgs", max_base_pg); | |
4380 | if (newmap) | |
4381 | f->dump_unsigned("new_max_osd_pgs", max_new_pg); | |
4382 | } else { | |
4383 | ss << "max osd." << max << " with " << max_base_pg; | |
4384 | if (newmap) | |
4385 | ss << " -> " << max_new_pg; | |
4386 | ss << " pgs (" << (float)max_base_pg / avg_pg; | |
4387 | if (newmap) | |
4388 | ss << " -> " << (float)max_new_pg / avg_pg; | |
4389 | ss << " * mean)\n"; | |
4390 | } | |
4391 | } | |
4392 | if (f) | |
4393 | f->close_section(); | |
4394 | if (out) | |
4395 | *out = ss.str(); | |
4396 | return 0; | |
4397 | } | |
4398 | ||
7c673cae FG |
4399 | bool OSDMap::try_pg_upmap( |
4400 | CephContext *cct, | |
4401 | pg_t pg, ///< pg to potentially remap | |
4402 | const set<int>& overfull, ///< osds we'd want to evacuate | |
4403 | const vector<int>& underfull, ///< osds to move to, in order of preference | |
92f5a8d4 | 4404 | const vector<int>& more_underfull, ///< more osds only slightly underfull |
7c673cae FG |
4405 | vector<int> *orig, |
4406 | vector<int> *out) ///< resulting alternative mapping | |
4407 | { | |
4408 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
4409 | if (!pool) | |
4410 | return false; | |
31f18b77 | 4411 | int rule = crush->find_rule(pool->get_crush_rule(), pool->get_type(), |
7c673cae FG |
4412 | pool->get_size()); |
4413 | if (rule < 0) | |
4414 | return false; | |
4415 | ||
7c673cae FG |
4416 | // make sure there is something there to remap |
4417 | bool any = false; | |
4418 | for (auto osd : *orig) { | |
4419 | if (overfull.count(osd)) { | |
4420 | any = true; | |
4421 | break; | |
4422 | } | |
4423 | } | |
4424 | if (!any) { | |
4425 | return false; | |
4426 | } | |
4427 | ||
4428 | int r = crush->try_remap_rule( | |
4429 | cct, | |
4430 | rule, | |
4431 | pool->get_size(), | |
4432 | overfull, underfull, | |
92f5a8d4 | 4433 | more_underfull, |
7c673cae FG |
4434 | *orig, |
4435 | out); | |
4436 | if (r < 0) | |
4437 | return false; | |
4438 | if (*out == *orig) | |
4439 | return false; | |
4440 | return true; | |
4441 | } | |
4442 | ||
4443 | int OSDMap::calc_pg_upmaps( | |
4444 | CephContext *cct, | |
92f5a8d4 | 4445 | uint32_t max_deviation, |
7c673cae | 4446 | int max, |
a8e16298 | 4447 | const set<int64_t>& only_pools, |
7c673cae FG |
4448 | OSDMap::Incremental *pending_inc) |
4449 | { | |
a8e16298 | 4450 | ldout(cct, 10) << __func__ << " pools " << only_pools << dendl; |
7c673cae | 4451 | OSDMap tmp; |
92f5a8d4 TL |
4452 | // Can't be less than 1 pg |
4453 | if (max_deviation < 1) | |
4454 | max_deviation = 1; | |
7c673cae FG |
4455 | tmp.deepish_copy_from(*this); |
4456 | int num_changed = 0; | |
a8e16298 TL |
4457 | map<int,set<pg_t>> pgs_by_osd; |
4458 | int total_pgs = 0; | |
4459 | float osd_weight_total = 0; | |
4460 | map<int,float> osd_weight; | |
4461 | for (auto& i : pools) { | |
4462 | if (!only_pools.empty() && !only_pools.count(i.first)) | |
4463 | continue; | |
4464 | for (unsigned ps = 0; ps < i.second.get_pg_num(); ++ps) { | |
4465 | pg_t pg(ps, i.first); | |
4466 | vector<int> up; | |
4467 | tmp.pg_to_up_acting_osds(pg, &up, nullptr, nullptr, nullptr); | |
4468 | ldout(cct, 20) << __func__ << " " << pg << " up " << up << dendl; | |
4469 | for (auto osd : up) { | |
4470 | if (osd != CRUSH_ITEM_NONE) | |
4471 | pgs_by_osd[osd].insert(pg); | |
7c673cae | 4472 | } |
a8e16298 TL |
4473 | } |
4474 | total_pgs += i.second.get_size() * i.second.get_pg_num(); | |
4475 | ||
4476 | map<int,float> pmap; | |
4477 | int ruleno = tmp.crush->find_rule(i.second.get_crush_rule(), | |
4478 | i.second.get_type(), | |
4479 | i.second.get_size()); | |
4480 | tmp.crush->get_rule_weight_osd_map(ruleno, &pmap); | |
4481 | ldout(cct,20) << __func__ << " pool " << i.first | |
4482 | << " ruleno " << ruleno | |
4483 | << " weight-map " << pmap | |
4484 | << dendl; | |
4485 | for (auto p : pmap) { | |
4486 | auto adjusted_weight = tmp.get_weightf(p.first) * p.second; | |
4487 | if (adjusted_weight == 0) { | |
4488 | continue; | |
31f18b77 | 4489 | } |
a8e16298 TL |
4490 | osd_weight[p.first] += adjusted_weight; |
4491 | osd_weight_total += adjusted_weight; | |
7c673cae | 4492 | } |
a8e16298 TL |
4493 | } |
4494 | for (auto& i : osd_weight) { | |
4495 | int pgs = 0; | |
4496 | auto p = pgs_by_osd.find(i.first); | |
4497 | if (p != pgs_by_osd.end()) | |
31f18b77 | 4498 | pgs = p->second.size(); |
a8e16298 | 4499 | else |
31f18b77 | 4500 | pgs_by_osd.emplace(i.first, set<pg_t>()); |
a8e16298 | 4501 | ldout(cct, 20) << " osd." << i.first << " weight " << i.second |
31f18b77 | 4502 | << " pgs " << pgs << dendl; |
a8e16298 TL |
4503 | } |
4504 | if (osd_weight_total == 0) { | |
4505 | lderr(cct) << __func__ << " abort due to osd_weight_total == 0" << dendl; | |
4506 | return 0; | |
4507 | } | |
4508 | float pgs_per_weight = total_pgs / osd_weight_total; | |
4509 | ldout(cct, 10) << " osd_weight_total " << osd_weight_total << dendl; | |
4510 | ldout(cct, 10) << " pgs_per_weight " << pgs_per_weight << dendl; | |
7c673cae | 4511 | |
a8e16298 TL |
4512 | if (max <= 0) { |
4513 | lderr(cct) << __func__ << " abort due to max <= 0" << dendl; | |
4514 | return 0; | |
4515 | } | |
a8e16298 TL |
4516 | float stddev = 0; |
4517 | map<int,float> osd_deviation; // osd, deviation(pgs) | |
4518 | multimap<float,int> deviation_osd; // deviation(pgs), osd | |
92f5a8d4 | 4519 | float cur_max_deviation = 0; |
a8e16298 TL |
4520 | for (auto& i : pgs_by_osd) { |
4521 | // make sure osd is still there (belongs to this crush-tree) | |
4522 | ceph_assert(osd_weight.count(i.first)); | |
4523 | float target = osd_weight[i.first] * pgs_per_weight; | |
4524 | float deviation = (float)i.second.size() - target; | |
4525 | ldout(cct, 20) << " osd." << i.first | |
4526 | << "\tpgs " << i.second.size() | |
4527 | << "\ttarget " << target | |
4528 | << "\tdeviation " << deviation | |
4529 | << dendl; | |
4530 | osd_deviation[i.first] = deviation; | |
4531 | deviation_osd.insert(make_pair(deviation, i.first)); | |
4532 | stddev += deviation * deviation; | |
92f5a8d4 TL |
4533 | if (fabsf(deviation) > cur_max_deviation) |
4534 | cur_max_deviation = fabsf(deviation); | |
a8e16298 | 4535 | } |
92f5a8d4 TL |
4536 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
4537 | if (cur_max_deviation <= max_deviation) { | |
a8e16298 TL |
4538 | ldout(cct, 10) << __func__ << " distribution is almost perfect" |
4539 | << dendl; | |
4540 | return 0; | |
4541 | } | |
4542 | bool skip_overfull = false; | |
4543 | auto aggressive = | |
11fdf7f2 | 4544 | cct->_conf.get_val<bool>("osd_calc_pg_upmaps_aggressively"); |
a8e16298 | 4545 | auto local_fallback_retries = |
11fdf7f2 | 4546 | cct->_conf.get_val<uint64_t>("osd_calc_pg_upmaps_local_fallback_retries"); |
a8e16298 | 4547 | while (max--) { |
92f5a8d4 | 4548 | ldout(cct, 30) << "Top of loop #" << max+1 << dendl; |
a8e16298 TL |
4549 | // build overfull and underfull |
4550 | set<int> overfull; | |
92f5a8d4 TL |
4551 | set<int> more_overfull; |
4552 | bool using_more_overfull = false; | |
a8e16298 | 4553 | vector<int> underfull; |
92f5a8d4 TL |
4554 | vector<int> more_underfull; |
4555 | for (auto i = deviation_osd.rbegin(); i != deviation_osd.rend(); i++) { | |
4556 | ldout(cct, 30) << " check " << i->first << " <= " << max_deviation << dendl; | |
4557 | if (i->first <= 0) | |
4558 | break; | |
4559 | if (i->first > max_deviation) { | |
4560 | ldout(cct, 30) << " add overfull osd." << i->second << dendl; | |
a8e16298 | 4561 | overfull.insert(i->second); |
92f5a8d4 TL |
4562 | } else { |
4563 | more_overfull.insert(i->second); | |
4564 | } | |
a8e16298 | 4565 | } |
7c673cae | 4566 | |
92f5a8d4 TL |
4567 | for (auto i = deviation_osd.begin(); i != deviation_osd.end(); i++) { |
4568 | ldout(cct, 30) << " check " << i->first << " >= " << -(int)max_deviation << dendl; | |
4569 | if (i->first >= 0) | |
a8e16298 | 4570 | break; |
92f5a8d4 TL |
4571 | if (i->first < -(int)max_deviation) { |
4572 | ldout(cct, 30) << " add underfull osd." << i->second << dendl; | |
4573 | underfull.push_back(i->second); | |
4574 | } else { | |
4575 | more_underfull.push_back(i->second); | |
4576 | } | |
7c673cae | 4577 | } |
92f5a8d4 TL |
4578 | if (underfull.empty() && overfull.empty()) { |
4579 | ldout(cct, 20) << __func__ << " failed to build overfull and underfull" << dendl; | |
7c673cae | 4580 | break; |
a8e16298 | 4581 | } |
92f5a8d4 TL |
4582 | if (overfull.empty() && !underfull.empty()) { |
4583 | ldout(cct, 20) << __func__ << " Using more_overfull since we still have underfull" << dendl; | |
4584 | overfull = more_overfull; | |
4585 | using_more_overfull = true; | |
4586 | } | |
7c673cae | 4587 | |
a8e16298 TL |
4588 | ldout(cct, 10) << " overfull " << overfull |
4589 | << " underfull " << underfull | |
4590 | << dendl; | |
4591 | set<pg_t> to_skip; | |
4592 | uint64_t local_fallback_retried = 0; | |
4593 | ||
4594 | retry: | |
4595 | ||
4596 | set<pg_t> to_unmap; | |
4597 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>> to_upmap; | |
4598 | auto temp_pgs_by_osd = pgs_by_osd; | |
4599 | // always start with fullest, break if we find any changes to make | |
7c673cae | 4600 | for (auto p = deviation_osd.rbegin(); p != deviation_osd.rend(); ++p) { |
92f5a8d4 | 4601 | if (skip_overfull && !underfull.empty()) { |
a8e16298 TL |
4602 | ldout(cct, 10) << " skipping overfull " << dendl; |
4603 | break; // fall through to check underfull | |
4604 | } | |
7c673cae | 4605 | int osd = p->second; |
31f18b77 | 4606 | float deviation = p->first; |
7c673cae | 4607 | float target = osd_weight[osd] * pgs_per_weight; |
92f5a8d4 TL |
4608 | ldout(cct, 10) << " Overfull search osd." << osd |
4609 | << " target " << target | |
4610 | << " deviation " << deviation | |
4611 | << dendl; | |
a8e16298 | 4612 | ceph_assert(target > 0); |
92f5a8d4 | 4613 | if (!using_more_overfull && deviation <= max_deviation) { |
7c673cae | 4614 | ldout(cct, 10) << " osd." << osd |
a8e16298 TL |
4615 | << " target " << target |
4616 | << " deviation " << deviation | |
92f5a8d4 | 4617 | << " < max deviation " << max_deviation |
a8e16298 | 4618 | << dendl; |
7c673cae FG |
4619 | break; |
4620 | } | |
7c673cae | 4621 | |
a8e16298 TL |
4622 | vector<pg_t> pgs; |
4623 | pgs.reserve(pgs_by_osd[osd].size()); | |
4624 | for (auto& pg : pgs_by_osd[osd]) { | |
4625 | if (to_skip.count(pg)) | |
4626 | continue; | |
4627 | pgs.push_back(pg); | |
4628 | } | |
4629 | if (aggressive) { | |
4630 | // shuffle PG list so they all get equal (in)attention | |
4631 | std::random_device rd; | |
4632 | std::default_random_engine rng{rd()}; | |
4633 | std::shuffle(pgs.begin(), pgs.end(), rng); | |
4634 | } | |
7c673cae FG |
4635 | // look for remaps we can un-remap |
4636 | for (auto pg : pgs) { | |
4637 | auto p = tmp.pg_upmap_items.find(pg); | |
a8e16298 TL |
4638 | if (p == tmp.pg_upmap_items.end()) |
4639 | continue; | |
4640 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
4641 | for (auto q : p->second) { | |
4642 | if (q.second == osd) { | |
4643 | ldout(cct, 10) << " will try dropping existing" | |
4644 | << " remapping pair " | |
4645 | << q.first << " -> " << q.second | |
4646 | << " which remapped " << pg | |
4647 | << " into overfull osd." << osd | |
4648 | << dendl; | |
4649 | temp_pgs_by_osd[q.second].erase(pg); | |
4650 | temp_pgs_by_osd[q.first].insert(pg); | |
4651 | } else { | |
4652 | new_upmap_items.push_back(q); | |
4653 | } | |
4654 | } | |
4655 | if (new_upmap_items.empty()) { | |
4656 | // drop whole item | |
4657 | ldout(cct, 10) << " existing pg_upmap_items " << p->second | |
4658 | << " remapped " << pg << " into overfull osd." << osd | |
4659 | << ", will try cancelling it entirely" | |
4660 | << dendl; | |
4661 | to_unmap.insert(pg); | |
4662 | goto test_change; | |
4663 | } else if (new_upmap_items.size() != p->second.size()) { | |
4664 | // drop single remapping pair, updating | |
4665 | ceph_assert(new_upmap_items.size() < p->second.size()); | |
4666 | ldout(cct, 10) << " existing pg_upmap_items " << p->second | |
4667 | << " remapped " << pg << " into overfull osd." << osd | |
4668 | << ", new_pg_upmap_items now " << new_upmap_items | |
4669 | << dendl; | |
4670 | to_upmap[pg] = new_upmap_items; | |
4671 | goto test_change; | |
4672 | } | |
4673 | } | |
7c673cae | 4674 | |
a8e16298 | 4675 | // try upmap |
7c673cae | 4676 | for (auto pg : pgs) { |
a8e16298 TL |
4677 | auto temp_it = tmp.pg_upmap.find(pg); |
4678 | if (temp_it != tmp.pg_upmap.end()) { | |
4679 | // leave pg_upmap alone | |
4680 | // it must be specified by admin since balancer does not | |
4681 | // support pg_upmap yet | |
4682 | ldout(cct, 10) << " " << pg << " already has pg_upmap " | |
4683 | << temp_it->second << ", skipping" | |
4684 | << dendl; | |
7c673cae FG |
4685 | continue; |
4686 | } | |
a8e16298 TL |
4687 | auto pg_pool_size = tmp.get_pg_pool_size(pg); |
4688 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
4689 | set<int> existing; | |
4690 | auto it = tmp.pg_upmap_items.find(pg); | |
4691 | if (it != tmp.pg_upmap_items.end() && | |
4692 | it->second.size() >= (size_t)pg_pool_size) { | |
4693 | ldout(cct, 10) << " " << pg << " already has full-size pg_upmap_items " | |
4694 | << it->second << ", skipping" | |
4695 | << dendl; | |
4696 | continue; | |
4697 | } else if (it != tmp.pg_upmap_items.end()) { | |
4698 | ldout(cct, 10) << " " << pg << " already has pg_upmap_items " | |
4699 | << it->second | |
4700 | << dendl; | |
4701 | new_upmap_items = it->second; | |
4702 | // build existing too (for dedup) | |
4703 | for (auto i : it->second) { | |
4704 | existing.insert(i.first); | |
4705 | existing.insert(i.second); | |
4706 | } | |
4707 | // fall through | |
4708 | // to see if we can append more remapping pairs | |
4709 | } | |
4710 | ldout(cct, 10) << " trying " << pg << dendl; | |
494da23a TL |
4711 | vector<int> raw, orig, out; |
4712 | tmp.pg_to_raw_upmap(pg, &raw, &orig); // including existing upmaps too | |
92f5a8d4 | 4713 | if (!try_pg_upmap(cct, pg, overfull, underfull, more_underfull, &orig, &out)) { |
7c673cae FG |
4714 | continue; |
4715 | } | |
a8e16298 | 4716 | ldout(cct, 10) << " " << pg << " " << orig << " -> " << out << dendl; |
7c673cae FG |
4717 | if (orig.size() != out.size()) { |
4718 | continue; | |
4719 | } | |
a8e16298 | 4720 | ceph_assert(orig != out); |
92f5a8d4 TL |
4721 | int pos = -1; |
4722 | float max_dev = 0; | |
7c673cae | 4723 | for (unsigned i = 0; i < out.size(); ++i) { |
a8e16298 TL |
4724 | if (orig[i] == out[i]) |
4725 | continue; // skip invalid remappings | |
4726 | if (existing.count(orig[i]) || existing.count(out[i])) | |
4727 | continue; // we want new remappings only! | |
92f5a8d4 TL |
4728 | if (osd_deviation[orig[i]] > max_dev) { |
4729 | max_dev = osd_deviation[orig[i]]; | |
4730 | pos = i; | |
4731 | ldout(cct, 30) << "Max osd." << orig[i] << " pos " << i << " dev " << osd_deviation[orig[i]] << dendl; | |
4732 | } | |
4733 | } | |
4734 | if (pos != -1) { | |
4735 | int i = pos; | |
a8e16298 TL |
4736 | ldout(cct, 10) << " will try adding new remapping pair " |
4737 | << orig[i] << " -> " << out[i] << " for " << pg | |
92f5a8d4 | 4738 | << (orig[i] != osd ? " NOT selected osd" : "") |
a8e16298 TL |
4739 | << dendl; |
4740 | existing.insert(orig[i]); | |
4741 | existing.insert(out[i]); | |
4742 | temp_pgs_by_osd[orig[i]].erase(pg); | |
4743 | temp_pgs_by_osd[out[i]].insert(pg); | |
4744 | ceph_assert(new_upmap_items.size() < (size_t)pg_pool_size); | |
4745 | new_upmap_items.push_back(make_pair(orig[i], out[i])); | |
4746 | // append new remapping pairs slowly | |
4747 | // This way we can make sure that each tiny change will | |
4748 | // definitely make distribution of PGs converging to | |
4749 | // the perfect status. | |
4750 | to_upmap[pg] = new_upmap_items; | |
4751 | goto test_change; | |
7c673cae | 4752 | } |
a8e16298 TL |
4753 | } |
4754 | } | |
7c673cae | 4755 | |
a8e16298 TL |
4756 | ceph_assert(!(to_unmap.size() || to_upmap.size())); |
4757 | ldout(cct, 10) << " failed to find any changes for overfull osds" | |
4758 | << dendl; | |
4759 | for (auto& p : deviation_osd) { | |
4760 | if (std::find(underfull.begin(), underfull.end(), p.second) == | |
4761 | underfull.end()) | |
4762 | break; | |
4763 | int osd = p.second; | |
4764 | float deviation = p.first; | |
4765 | float target = osd_weight[osd] * pgs_per_weight; | |
4766 | ceph_assert(target > 0); | |
92f5a8d4 TL |
4767 | if (fabsf(deviation) < max_deviation) { |
4768 | // respect max_deviation too | |
a8e16298 TL |
4769 | ldout(cct, 10) << " osd." << osd |
4770 | << " target " << target | |
4771 | << " deviation " << deviation | |
92f5a8d4 TL |
4772 | << " -> absolute " << fabsf(deviation) |
4773 | << " < max " << max_deviation | |
a8e16298 TL |
4774 | << dendl; |
4775 | break; | |
4776 | } | |
4777 | // look for remaps we can un-remap | |
4778 | vector<pair<pg_t, | |
4779 | mempool::osdmap::vector<pair<int32_t,int32_t>>>> candidates; | |
4780 | candidates.reserve(tmp.pg_upmap_items.size()); | |
4781 | for (auto& i : tmp.pg_upmap_items) { | |
4782 | if (to_skip.count(i.first)) | |
4783 | continue; | |
4784 | if (!only_pools.empty() && !only_pools.count(i.first.pool())) | |
4785 | continue; | |
4786 | candidates.push_back(make_pair(i.first, i.second)); | |
4787 | } | |
4788 | if (aggressive) { | |
4789 | // shuffle candidates so they all get equal (in)attention | |
4790 | std::random_device rd; | |
4791 | std::default_random_engine rng{rd()}; | |
4792 | std::shuffle(candidates.begin(), candidates.end(), rng); | |
4793 | } | |
4794 | for (auto& i : candidates) { | |
4795 | auto pg = i.first; | |
4796 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
4797 | for (auto& j : i.second) { | |
4798 | if (j.first == osd) { | |
4799 | ldout(cct, 10) << " will try dropping existing" | |
4800 | << " remapping pair " | |
4801 | << j.first << " -> " << j.second | |
4802 | << " which remapped " << pg | |
4803 | << " out from underfull osd." << osd | |
4804 | << dendl; | |
4805 | temp_pgs_by_osd[j.second].erase(pg); | |
4806 | temp_pgs_by_osd[j.first].insert(pg); | |
4807 | } else { | |
4808 | new_upmap_items.push_back(j); | |
4809 | } | |
4810 | } | |
4811 | if (new_upmap_items.empty()) { | |
4812 | // drop whole item | |
4813 | ldout(cct, 10) << " existing pg_upmap_items " << i.second | |
4814 | << " remapped " << pg | |
4815 | << " out from underfull osd." << osd | |
4816 | << ", will try cancelling it entirely" | |
4817 | << dendl; | |
4818 | to_unmap.insert(pg); | |
4819 | goto test_change; | |
4820 | } else if (new_upmap_items.size() != i.second.size()) { | |
4821 | // drop single remapping pair, updating | |
4822 | ceph_assert(new_upmap_items.size() < i.second.size()); | |
4823 | ldout(cct, 10) << " existing pg_upmap_items " << i.second | |
4824 | << " remapped " << pg | |
4825 | << " out from underfull osd." << osd | |
4826 | << ", new_pg_upmap_items now " << new_upmap_items | |
4827 | << dendl; | |
4828 | to_upmap[pg] = new_upmap_items; | |
4829 | goto test_change; | |
4830 | } | |
4831 | } | |
7c673cae | 4832 | } |
a8e16298 TL |
4833 | |
4834 | ceph_assert(!(to_unmap.size() || to_upmap.size())); | |
4835 | ldout(cct, 10) << " failed to find any changes for underfull osds" | |
4836 | << dendl; | |
4837 | if (!aggressive) { | |
4838 | ldout(cct, 10) << " break due to aggressive mode not enabled" << dendl; | |
4839 | break; | |
4840 | } else if (!skip_overfull) { | |
4841 | // safe to quit because below here we know | |
4842 | // we've done checking both overfull and underfull osds.. | |
4843 | ldout(cct, 10) << " break due to not being able to find any" | |
4844 | << " further optimizations" | |
4845 | << dendl; | |
7c673cae FG |
4846 | break; |
4847 | } | |
a8e16298 TL |
4848 | // restart with fullest and do exhaustive searching |
4849 | skip_overfull = false; | |
4850 | continue; | |
4851 | ||
4852 | test_change: | |
4853 | ||
4854 | // test change, apply if change is good | |
4855 | ceph_assert(to_unmap.size() || to_upmap.size()); | |
4856 | float new_stddev = 0; | |
4857 | map<int,float> temp_osd_deviation; | |
4858 | multimap<float,int> temp_deviation_osd; | |
92f5a8d4 | 4859 | float cur_max_deviation = 0; |
a8e16298 TL |
4860 | for (auto& i : temp_pgs_by_osd) { |
4861 | // make sure osd is still there (belongs to this crush-tree) | |
4862 | ceph_assert(osd_weight.count(i.first)); | |
4863 | float target = osd_weight[i.first] * pgs_per_weight; | |
4864 | float deviation = (float)i.second.size() - target; | |
4865 | ldout(cct, 20) << " osd." << i.first | |
4866 | << "\tpgs " << i.second.size() | |
4867 | << "\ttarget " << target | |
4868 | << "\tdeviation " << deviation | |
4869 | << dendl; | |
4870 | temp_osd_deviation[i.first] = deviation; | |
4871 | temp_deviation_osd.insert(make_pair(deviation, i.first)); | |
92f5a8d4 TL |
4872 | new_stddev += deviation * deviation; |
4873 | if (fabsf(deviation) > cur_max_deviation) | |
4874 | cur_max_deviation = fabsf(deviation); | |
a8e16298 TL |
4875 | } |
4876 | ldout(cct, 10) << " stddev " << stddev << " -> " << new_stddev << dendl; | |
4877 | if (new_stddev >= stddev) { | |
4878 | if (!aggressive) { | |
4879 | ldout(cct, 10) << " break because stddev is not decreasing" | |
4880 | << " and aggressive mode is not enabled" | |
4881 | << dendl; | |
4882 | break; | |
4883 | } | |
4884 | local_fallback_retried++; | |
4885 | if (local_fallback_retried >= local_fallback_retries) { | |
4886 | // does not make progress | |
4887 | // flip *skip_overfull* so both overfull and underfull | |
4888 | // get equal (in)attention | |
4889 | skip_overfull = !skip_overfull; | |
4890 | ldout(cct, 10) << " hit local_fallback_retries " | |
4891 | << local_fallback_retries | |
4892 | << dendl; | |
4893 | continue; | |
4894 | } | |
4895 | for (auto& i : to_unmap) | |
4896 | to_skip.insert(i); | |
4897 | for (auto& i : to_upmap) | |
4898 | to_skip.insert(i.first); | |
4899 | ldout(cct, 20) << " local_fallback_retried " << local_fallback_retried | |
4900 | << " to_skip " << to_skip | |
4901 | << dendl; | |
4902 | goto retry; | |
4903 | } | |
4904 | ||
4905 | // ready to go | |
4906 | ceph_assert(new_stddev < stddev); | |
4907 | stddev = new_stddev; | |
4908 | pgs_by_osd = temp_pgs_by_osd; | |
4909 | osd_deviation = temp_osd_deviation; | |
4910 | deviation_osd = temp_deviation_osd; | |
4911 | for (auto& i : to_unmap) { | |
4912 | ldout(cct, 10) << " unmap pg " << i << dendl; | |
4913 | ceph_assert(tmp.pg_upmap_items.count(i)); | |
4914 | tmp.pg_upmap_items.erase(i); | |
4915 | pending_inc->old_pg_upmap_items.insert(i); | |
4916 | ++num_changed; | |
4917 | } | |
4918 | for (auto& i : to_upmap) { | |
4919 | ldout(cct, 10) << " upmap pg " << i.first | |
4920 | << " new pg_upmap_items " << i.second | |
4921 | << dendl; | |
4922 | tmp.pg_upmap_items[i.first] = i.second; | |
4923 | pending_inc->new_pg_upmap_items[i.first] = i.second; | |
4924 | ++num_changed; | |
4925 | } | |
92f5a8d4 TL |
4926 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
4927 | if (cur_max_deviation <= max_deviation) { | |
4928 | ldout(cct, 10) << __func__ << " Optimization plan is almost perfect" | |
4929 | << dendl; | |
4930 | break; | |
4931 | } | |
7c673cae | 4932 | } |
a8e16298 | 4933 | ldout(cct, 10) << " num_changed = " << num_changed << dendl; |
7c673cae FG |
4934 | return num_changed; |
4935 | } | |
31f18b77 FG |
4936 | |
4937 | int OSDMap::get_osds_by_bucket_name(const string &name, set<int> *osds) const | |
4938 | { | |
4939 | return crush->get_leaves(name, osds); | |
4940 | } | |
4941 | ||
3efd9988 FG |
4942 | // get pools whose crush rules might reference the given osd |
4943 | void OSDMap::get_pool_ids_by_osd(CephContext *cct, | |
4944 | int osd, | |
4945 | set<int64_t> *pool_ids) const | |
4946 | { | |
11fdf7f2 | 4947 | ceph_assert(pool_ids); |
3efd9988 FG |
4948 | set<int> raw_rules; |
4949 | int r = crush->get_rules_by_osd(osd, &raw_rules); | |
4950 | if (r < 0) { | |
4951 | lderr(cct) << __func__ << " get_rules_by_osd failed: " << cpp_strerror(r) | |
4952 | << dendl; | |
11fdf7f2 | 4953 | ceph_assert(r >= 0); |
3efd9988 FG |
4954 | } |
4955 | set<int> rules; | |
4956 | for (auto &i: raw_rules) { | |
4957 | // exclude any dead rule | |
4958 | if (crush_rule_in_use(i)) { | |
4959 | rules.insert(i); | |
4960 | } | |
4961 | } | |
4962 | for (auto &r: rules) { | |
4963 | get_pool_ids_by_rule(r, pool_ids); | |
4964 | } | |
4965 | } | |
4966 | ||
31f18b77 FG |
4967 | template <typename F> |
4968 | class OSDUtilizationDumper : public CrushTreeDumper::Dumper<F> { | |
4969 | public: | |
4970 | typedef CrushTreeDumper::Dumper<F> Parent; | |
4971 | ||
4972 | OSDUtilizationDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
11fdf7f2 TL |
4973 | const PGMap& pgmap_, bool tree_, |
4974 | const string& class_name_, | |
4975 | const string& item_name_) : | |
c07f9fc5 | 4976 | Parent(crush, osdmap_->get_pool_names()), |
31f18b77 | 4977 | osdmap(osdmap_), |
11fdf7f2 | 4978 | pgmap(pgmap_), |
31f18b77 | 4979 | tree(tree_), |
11fdf7f2 TL |
4980 | class_name(class_name_), |
4981 | item_name(item_name_), | |
31f18b77 FG |
4982 | min_var(-1), |
4983 | max_var(-1), | |
4984 | stddev(0), | |
4985 | sum(0) { | |
11fdf7f2 TL |
4986 | if (osdmap->crush->name_exists(item_name)) { |
4987 | // filter out items we are allowed to dump | |
4988 | auto item_id = osdmap->crush->get_item_id(item_name); | |
4989 | allowed.insert(item_id); | |
4990 | osdmap->crush->get_all_children(item_id, &allowed); | |
4991 | } | |
4992 | average_util = average_utilization(); | |
31f18b77 FG |
4993 | } |
4994 | ||
4995 | protected: | |
11fdf7f2 TL |
4996 | |
4997 | bool should_dump(int id) const { | |
4998 | if (!allowed.empty() && !allowed.count(id)) // filter by name | |
4999 | return false; | |
5000 | if (id >= 0 && !class_name.empty()) { | |
5001 | const char* item_class_name = osdmap->crush->get_item_class(id); | |
5002 | if (!item_class_name || // not bound to a class yet | |
5003 | item_class_name != class_name) // or already bound to | |
5004 | // a different class | |
5005 | return false; | |
5006 | } | |
5007 | return true; | |
5008 | } | |
5009 | ||
5010 | set<int> get_dumped_osds() { | |
5011 | if (class_name.empty() && item_name.empty()) { | |
5012 | // old way, all | |
5013 | return {}; | |
5014 | } | |
5015 | return dumped_osds; | |
5016 | } | |
5017 | ||
31f18b77 FG |
5018 | void dump_stray(F *f) { |
5019 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
5020 | if (osdmap->exists(i) && !this->is_touched(i)) | |
c07f9fc5 | 5021 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
31f18b77 FG |
5022 | } |
5023 | } | |
5024 | ||
5025 | void dump_item(const CrushTreeDumper::Item &qi, F *f) override { | |
5026 | if (!tree && qi.is_bucket()) | |
5027 | return; | |
11fdf7f2 TL |
5028 | if (!should_dump(qi.id)) |
5029 | return; | |
31f18b77 | 5030 | |
11fdf7f2 TL |
5031 | if (!qi.is_bucket()) |
5032 | dumped_osds.insert(qi.id); | |
31f18b77 | 5033 | float reweight = qi.is_bucket() ? -1 : osdmap->get_weightf(qi.id); |
11fdf7f2 TL |
5034 | int64_t kb = 0, kb_used = 0, kb_used_data = 0, kb_used_omap = 0, |
5035 | kb_used_meta = 0, kb_avail = 0; | |
31f18b77 | 5036 | double util = 0; |
11fdf7f2 TL |
5037 | if (get_bucket_utilization(qi.id, &kb, &kb_used, &kb_used_data, |
5038 | &kb_used_omap, &kb_used_meta, &kb_avail)) | |
31f18b77 FG |
5039 | if (kb_used && kb) |
5040 | util = 100.0 * (double)kb_used / (double)kb; | |
5041 | ||
5042 | double var = 1.0; | |
5043 | if (average_util) | |
5044 | var = util / average_util; | |
5045 | ||
11fdf7f2 | 5046 | size_t num_pgs = qi.is_bucket() ? 0 : pgmap.get_num_pg_by_osd(qi.id); |
31f18b77 | 5047 | |
11fdf7f2 TL |
5048 | dump_item(qi, reweight, kb, kb_used, |
5049 | kb_used_data, kb_used_omap, kb_used_meta, | |
5050 | kb_avail, util, var, num_pgs, f); | |
31f18b77 FG |
5051 | |
5052 | if (!qi.is_bucket() && reweight > 0) { | |
5053 | if (min_var < 0 || var < min_var) | |
5054 | min_var = var; | |
5055 | if (max_var < 0 || var > max_var) | |
5056 | max_var = var; | |
5057 | ||
5058 | double dev = util - average_util; | |
5059 | dev *= dev; | |
5060 | stddev += reweight * dev; | |
5061 | sum += reweight; | |
5062 | } | |
5063 | } | |
5064 | ||
5065 | virtual void dump_item(const CrushTreeDumper::Item &qi, | |
5066 | float &reweight, | |
5067 | int64_t kb, | |
5068 | int64_t kb_used, | |
11fdf7f2 TL |
5069 | int64_t kb_used_data, |
5070 | int64_t kb_used_omap, | |
5071 | int64_t kb_used_meta, | |
31f18b77 FG |
5072 | int64_t kb_avail, |
5073 | double& util, | |
5074 | double& var, | |
5075 | const size_t num_pgs, | |
5076 | F *f) = 0; | |
5077 | ||
5078 | double dev() { | |
5079 | return sum > 0 ? sqrt(stddev / sum) : 0; | |
5080 | } | |
5081 | ||
5082 | double average_utilization() { | |
5083 | int64_t kb = 0, kb_used = 0; | |
5084 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
11fdf7f2 TL |
5085 | if (!osdmap->exists(i) || |
5086 | osdmap->get_weight(i) == 0 || | |
5087 | !should_dump(i)) | |
31f18b77 | 5088 | continue; |
11fdf7f2 TL |
5089 | int64_t kb_i, kb_used_i, kb_used_data_i, kb_used_omap_i, kb_used_meta_i, |
5090 | kb_avail_i; | |
5091 | if (get_osd_utilization(i, &kb_i, &kb_used_i, &kb_used_data_i, | |
5092 | &kb_used_omap_i, &kb_used_meta_i, &kb_avail_i)) { | |
31f18b77 FG |
5093 | kb += kb_i; |
5094 | kb_used += kb_used_i; | |
5095 | } | |
5096 | } | |
5097 | return kb > 0 ? 100.0 * (double)kb_used / (double)kb : 0; | |
5098 | } | |
5099 | ||
5100 | bool get_osd_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5101 | int64_t* kb_used_data, |
5102 | int64_t* kb_used_omap, | |
5103 | int64_t* kb_used_meta, | |
31f18b77 | 5104 | int64_t* kb_avail) const { |
11fdf7f2 | 5105 | const osd_stat_t *p = pgmap.get_osd_stat(id); |
31f18b77 | 5106 | if (!p) return false; |
11fdf7f2 TL |
5107 | *kb = p->statfs.kb(); |
5108 | *kb_used = p->statfs.kb_used_raw(); | |
5109 | *kb_used_data = p->statfs.kb_used_data(); | |
5110 | *kb_used_omap = p->statfs.kb_used_omap(); | |
5111 | *kb_used_meta = p->statfs.kb_used_internal_metadata(); | |
5112 | *kb_avail = p->statfs.kb_avail(); | |
5113 | ||
31f18b77 FG |
5114 | return *kb > 0; |
5115 | } | |
5116 | ||
5117 | bool get_bucket_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5118 | int64_t* kb_used_data, |
5119 | int64_t* kb_used_omap, | |
5120 | int64_t* kb_used_meta, | |
31f18b77 FG |
5121 | int64_t* kb_avail) const { |
5122 | if (id >= 0) { | |
11fdf7f2 | 5123 | if (osdmap->is_out(id) || !should_dump(id)) { |
31f18b77 FG |
5124 | *kb = 0; |
5125 | *kb_used = 0; | |
11fdf7f2 TL |
5126 | *kb_used_data = 0; |
5127 | *kb_used_omap = 0; | |
5128 | *kb_used_meta = 0; | |
31f18b77 FG |
5129 | *kb_avail = 0; |
5130 | return true; | |
5131 | } | |
11fdf7f2 TL |
5132 | return get_osd_utilization(id, kb, kb_used, kb_used_data, |
5133 | kb_used_omap, kb_used_meta, kb_avail); | |
31f18b77 FG |
5134 | } |
5135 | ||
5136 | *kb = 0; | |
5137 | *kb_used = 0; | |
11fdf7f2 TL |
5138 | *kb_used_data = 0; |
5139 | *kb_used_omap = 0; | |
5140 | *kb_used_meta = 0; | |
31f18b77 FG |
5141 | *kb_avail = 0; |
5142 | ||
5143 | for (int k = osdmap->crush->get_bucket_size(id) - 1; k >= 0; k--) { | |
5144 | int item = osdmap->crush->get_bucket_item(id, k); | |
11fdf7f2 TL |
5145 | int64_t kb_i = 0, kb_used_i = 0, kb_used_data_i = 0, |
5146 | kb_used_omap_i = 0, kb_used_meta_i = 0, kb_avail_i = 0; | |
5147 | if (!get_bucket_utilization(item, &kb_i, &kb_used_i, | |
5148 | &kb_used_data_i, &kb_used_omap_i, | |
5149 | &kb_used_meta_i, &kb_avail_i)) | |
31f18b77 FG |
5150 | return false; |
5151 | *kb += kb_i; | |
5152 | *kb_used += kb_used_i; | |
11fdf7f2 TL |
5153 | *kb_used_data += kb_used_data_i; |
5154 | *kb_used_omap += kb_used_omap_i; | |
5155 | *kb_used_meta += kb_used_meta_i; | |
31f18b77 FG |
5156 | *kb_avail += kb_avail_i; |
5157 | } | |
5158 | return *kb > 0; | |
5159 | } | |
5160 | ||
5161 | protected: | |
5162 | const OSDMap *osdmap; | |
11fdf7f2 | 5163 | const PGMap& pgmap; |
31f18b77 | 5164 | bool tree; |
11fdf7f2 TL |
5165 | const string class_name; |
5166 | const string item_name; | |
31f18b77 FG |
5167 | double average_util; |
5168 | double min_var; | |
5169 | double max_var; | |
5170 | double stddev; | |
5171 | double sum; | |
11fdf7f2 TL |
5172 | set<int> allowed; |
5173 | set<int> dumped_osds; | |
31f18b77 FG |
5174 | }; |
5175 | ||
5176 | ||
5177 | class OSDUtilizationPlainDumper : public OSDUtilizationDumper<TextTable> { | |
5178 | public: | |
5179 | typedef OSDUtilizationDumper<TextTable> Parent; | |
5180 | ||
5181 | OSDUtilizationPlainDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 TL |
5182 | const PGMap& pgmap, bool tree, |
5183 | const string& class_name, | |
5184 | const string& item_name) : | |
5185 | Parent(crush, osdmap, pgmap, tree, class_name, item_name) {} | |
31f18b77 FG |
5186 | |
5187 | void dump(TextTable *tbl) { | |
5188 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 5189 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5190 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
5191 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); | |
5192 | tbl->define_column("SIZE", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 TL |
5193 | tbl->define_column("RAW USE", TextTable::LEFT, TextTable::RIGHT); |
5194 | tbl->define_column("DATA", TextTable::LEFT, TextTable::RIGHT); | |
5195 | tbl->define_column("OMAP", TextTable::LEFT, TextTable::RIGHT); | |
5196 | tbl->define_column("META", TextTable::LEFT, TextTable::RIGHT); | |
31f18b77 FG |
5197 | tbl->define_column("AVAIL", TextTable::LEFT, TextTable::RIGHT); |
5198 | tbl->define_column("%USE", TextTable::LEFT, TextTable::RIGHT); | |
5199 | tbl->define_column("VAR", TextTable::LEFT, TextTable::RIGHT); | |
5200 | tbl->define_column("PGS", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 | 5201 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5202 | if (tree) |
5203 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
5204 | ||
5205 | Parent::dump(tbl); | |
5206 | ||
5207 | dump_stray(tbl); | |
5208 | ||
11fdf7f2 | 5209 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
224ce89b WB |
5210 | *tbl << "" |
5211 | << "" | |
5212 | << "" << "TOTAL" | |
11fdf7f2 TL |
5213 | << byte_u_t(sum.statfs.total) |
5214 | << byte_u_t(sum.statfs.get_used_raw()) | |
5215 | << byte_u_t(sum.statfs.allocated) | |
5216 | << byte_u_t(sum.statfs.omap_allocated) | |
5217 | << byte_u_t(sum.statfs.internal_metadata) | |
5218 | << byte_u_t(sum.statfs.available) | |
31f18b77 FG |
5219 | << lowprecision_t(average_util) |
5220 | << "" | |
5221 | << TextTable::endrow; | |
5222 | } | |
5223 | ||
5224 | protected: | |
5225 | struct lowprecision_t { | |
5226 | float v; | |
5227 | explicit lowprecision_t(float _v) : v(_v) {} | |
5228 | }; | |
5229 | friend std::ostream &operator<<(ostream& out, const lowprecision_t& v); | |
5230 | ||
5231 | using OSDUtilizationDumper<TextTable>::dump_item; | |
5232 | void dump_item(const CrushTreeDumper::Item &qi, | |
5233 | float &reweight, | |
5234 | int64_t kb, | |
5235 | int64_t kb_used, | |
11fdf7f2 TL |
5236 | int64_t kb_used_data, |
5237 | int64_t kb_used_omap, | |
5238 | int64_t kb_used_meta, | |
31f18b77 FG |
5239 | int64_t kb_avail, |
5240 | double& util, | |
5241 | double& var, | |
5242 | const size_t num_pgs, | |
5243 | TextTable *tbl) override { | |
224ce89b WB |
5244 | const char *c = crush->get_item_class(qi.id); |
5245 | if (!c) | |
5246 | c = ""; | |
31f18b77 | 5247 | *tbl << qi.id |
224ce89b | 5248 | << c |
31f18b77 FG |
5249 | << weightf_t(qi.weight) |
5250 | << weightf_t(reweight) | |
1adf2230 AA |
5251 | << byte_u_t(kb << 10) |
5252 | << byte_u_t(kb_used << 10) | |
11fdf7f2 TL |
5253 | << byte_u_t(kb_used_data << 10) |
5254 | << byte_u_t(kb_used_omap << 10) | |
5255 | << byte_u_t(kb_used_meta << 10) | |
1adf2230 | 5256 | << byte_u_t(kb_avail << 10) |
31f18b77 FG |
5257 | << lowprecision_t(util) |
5258 | << lowprecision_t(var); | |
5259 | ||
5260 | if (qi.is_bucket()) { | |
5261 | *tbl << "-"; | |
11fdf7f2 | 5262 | *tbl << ""; |
31f18b77 FG |
5263 | } else { |
5264 | *tbl << num_pgs; | |
11fdf7f2 TL |
5265 | if (osdmap->is_up(qi.id)) { |
5266 | *tbl << "up"; | |
5267 | } else if (osdmap->is_destroyed(qi.id)) { | |
5268 | *tbl << "destroyed"; | |
5269 | } else { | |
5270 | *tbl << "down"; | |
5271 | } | |
31f18b77 FG |
5272 | } |
5273 | ||
5274 | if (tree) { | |
5275 | ostringstream name; | |
5276 | for (int k = 0; k < qi.depth; k++) | |
5277 | name << " "; | |
5278 | if (qi.is_bucket()) { | |
5279 | int type = crush->get_bucket_type(qi.id); | |
5280 | name << crush->get_type_name(type) << " " | |
5281 | << crush->get_item_name(qi.id); | |
5282 | } else { | |
5283 | name << "osd." << qi.id; | |
5284 | } | |
5285 | *tbl << name.str(); | |
5286 | } | |
5287 | ||
5288 | *tbl << TextTable::endrow; | |
5289 | } | |
5290 | ||
5291 | public: | |
5292 | string summary() { | |
5293 | ostringstream out; | |
5294 | out << "MIN/MAX VAR: " << lowprecision_t(min_var) | |
5295 | << "/" << lowprecision_t(max_var) << " " | |
5296 | << "STDDEV: " << lowprecision_t(dev()); | |
5297 | return out.str(); | |
5298 | } | |
5299 | }; | |
5300 | ||
5301 | ostream& operator<<(ostream& out, | |
5302 | const OSDUtilizationPlainDumper::lowprecision_t& v) | |
5303 | { | |
5304 | if (v.v < -0.01) { | |
5305 | return out << "-"; | |
5306 | } else if (v.v < 0.001) { | |
5307 | return out << "0"; | |
5308 | } else { | |
5309 | std::streamsize p = out.precision(); | |
5310 | return out << std::fixed << std::setprecision(2) << v.v << std::setprecision(p); | |
5311 | } | |
5312 | } | |
5313 | ||
5314 | class OSDUtilizationFormatDumper : public OSDUtilizationDumper<Formatter> { | |
5315 | public: | |
5316 | typedef OSDUtilizationDumper<Formatter> Parent; | |
5317 | ||
5318 | OSDUtilizationFormatDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 TL |
5319 | const PGMap& pgmap, bool tree, |
5320 | const string& class_name, | |
5321 | const string& item_name) : | |
5322 | Parent(crush, osdmap, pgmap, tree, class_name, item_name) {} | |
31f18b77 FG |
5323 | |
5324 | void dump(Formatter *f) { | |
5325 | f->open_array_section("nodes"); | |
5326 | Parent::dump(f); | |
5327 | f->close_section(); | |
5328 | ||
5329 | f->open_array_section("stray"); | |
5330 | dump_stray(f); | |
5331 | f->close_section(); | |
5332 | } | |
5333 | ||
5334 | protected: | |
5335 | using OSDUtilizationDumper<Formatter>::dump_item; | |
5336 | void dump_item(const CrushTreeDumper::Item &qi, | |
11fdf7f2 TL |
5337 | float &reweight, |
5338 | int64_t kb, | |
5339 | int64_t kb_used, | |
5340 | int64_t kb_used_data, | |
5341 | int64_t kb_used_omap, | |
5342 | int64_t kb_used_meta, | |
5343 | int64_t kb_avail, | |
5344 | double& util, | |
5345 | double& var, | |
5346 | const size_t num_pgs, | |
5347 | Formatter *f) override { | |
31f18b77 | 5348 | f->open_object_section("item"); |
c07f9fc5 | 5349 | CrushTreeDumper::dump_item_fields(crush, weight_set_names, qi, f); |
31f18b77 FG |
5350 | f->dump_float("reweight", reweight); |
5351 | f->dump_int("kb", kb); | |
5352 | f->dump_int("kb_used", kb_used); | |
11fdf7f2 TL |
5353 | f->dump_int("kb_used_data", kb_used_data); |
5354 | f->dump_int("kb_used_omap", kb_used_omap); | |
5355 | f->dump_int("kb_used_meta", kb_used_meta); | |
31f18b77 FG |
5356 | f->dump_int("kb_avail", kb_avail); |
5357 | f->dump_float("utilization", util); | |
5358 | f->dump_float("var", var); | |
5359 | f->dump_unsigned("pgs", num_pgs); | |
11fdf7f2 TL |
5360 | if (!qi.is_bucket()) { |
5361 | if (osdmap->is_up(qi.id)) { | |
5362 | f->dump_string("status", "up"); | |
5363 | } else if (osdmap->is_destroyed(qi.id)) { | |
5364 | f->dump_string("status", "destroyed"); | |
5365 | } else { | |
5366 | f->dump_string("status", "down"); | |
5367 | } | |
5368 | } | |
31f18b77 FG |
5369 | CrushTreeDumper::dump_bucket_children(crush, qi, f); |
5370 | f->close_section(); | |
5371 | } | |
5372 | ||
5373 | public: | |
5374 | void summary(Formatter *f) { | |
5375 | f->open_object_section("summary"); | |
11fdf7f2 TL |
5376 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
5377 | auto& s = sum.statfs; | |
5378 | ||
5379 | f->dump_int("total_kb", s.kb()); | |
5380 | f->dump_int("total_kb_used", s.kb_used_raw()); | |
5381 | f->dump_int("total_kb_used_data", s.kb_used_data()); | |
5382 | f->dump_int("total_kb_used_omap", s.kb_used_omap()); | |
5383 | f->dump_int("total_kb_used_meta", s.kb_used_internal_metadata()); | |
5384 | f->dump_int("total_kb_avail", s.kb_avail()); | |
31f18b77 FG |
5385 | f->dump_float("average_utilization", average_util); |
5386 | f->dump_float("min_var", min_var); | |
5387 | f->dump_float("max_var", max_var); | |
5388 | f->dump_float("dev", dev()); | |
5389 | f->close_section(); | |
5390 | } | |
5391 | }; | |
5392 | ||
5393 | void print_osd_utilization(const OSDMap& osdmap, | |
11fdf7f2 TL |
5394 | const PGMap& pgmap, |
5395 | ostream& out, | |
5396 | Formatter *f, | |
5397 | bool tree, | |
5398 | const string& class_name, | |
5399 | const string& item_name) | |
31f18b77 FG |
5400 | { |
5401 | const CrushWrapper *crush = osdmap.crush.get(); | |
5402 | if (f) { | |
5403 | f->open_object_section("df"); | |
11fdf7f2 TL |
5404 | OSDUtilizationFormatDumper d(crush, &osdmap, pgmap, tree, |
5405 | class_name, item_name); | |
31f18b77 FG |
5406 | d.dump(f); |
5407 | d.summary(f); | |
5408 | f->close_section(); | |
5409 | f->flush(out); | |
5410 | } else { | |
11fdf7f2 TL |
5411 | OSDUtilizationPlainDumper d(crush, &osdmap, pgmap, tree, |
5412 | class_name, item_name); | |
31f18b77 FG |
5413 | TextTable tbl; |
5414 | d.dump(&tbl); | |
5415 | out << tbl << d.summary() << "\n"; | |
5416 | } | |
5417 | } | |
224ce89b | 5418 | |
92f5a8d4 TL |
5419 | void OSDMap::check_health(CephContext *cct, |
5420 | health_check_map_t *checks) const | |
224ce89b WB |
5421 | { |
5422 | int num_osds = get_num_osds(); | |
5423 | ||
5424 | // OSD_DOWN | |
5425 | // OSD_$subtree_DOWN | |
5426 | // OSD_ORPHAN | |
5427 | if (num_osds >= 0) { | |
5428 | int num_in_osds = 0; | |
5429 | int num_down_in_osds = 0; | |
5430 | set<int> osds; | |
5431 | set<int> down_in_osds; | |
5432 | set<int> up_in_osds; | |
5433 | set<int> subtree_up; | |
5434 | unordered_map<int, set<int> > subtree_type_down; | |
5435 | unordered_map<int, int> num_osds_subtree; | |
5436 | int max_type = crush->get_max_type_id(); | |
5437 | ||
5438 | for (int i = 0; i < get_max_osd(); i++) { | |
5439 | if (!exists(i)) { | |
5440 | if (crush->item_exists(i)) { | |
5441 | osds.insert(i); | |
5442 | } | |
5443 | continue; | |
5444 | } | |
5445 | if (is_out(i)) | |
5446 | continue; | |
5447 | ++num_in_osds; | |
5448 | if (down_in_osds.count(i) || up_in_osds.count(i)) | |
5449 | continue; | |
5450 | if (!is_up(i)) { | |
5451 | down_in_osds.insert(i); | |
5452 | int parent_id = 0; | |
5453 | int current = i; | |
5454 | for (int type = 0; type <= max_type; type++) { | |
5455 | if (!crush->get_type_name(type)) | |
5456 | continue; | |
5457 | int r = crush->get_immediate_parent_id(current, &parent_id); | |
5458 | if (r == -ENOENT) | |
5459 | break; | |
5460 | // break early if this parent is already marked as up | |
5461 | if (subtree_up.count(parent_id)) | |
5462 | break; | |
5463 | type = crush->get_bucket_type(parent_id); | |
5464 | if (!subtree_type_is_down( | |
92f5a8d4 | 5465 | cct, parent_id, type, |
224ce89b WB |
5466 | &down_in_osds, &up_in_osds, &subtree_up, &subtree_type_down)) |
5467 | break; | |
5468 | current = parent_id; | |
5469 | } | |
5470 | } | |
5471 | } | |
5472 | ||
5473 | // calculate the number of down osds in each down subtree and | |
5474 | // store it in num_osds_subtree | |
5475 | for (int type = 1; type <= max_type; type++) { | |
5476 | if (!crush->get_type_name(type)) | |
5477 | continue; | |
5478 | for (auto j = subtree_type_down[type].begin(); | |
5479 | j != subtree_type_down[type].end(); | |
5480 | ++j) { | |
5481 | list<int> children; | |
5482 | int num = 0; | |
5483 | int num_children = crush->get_children(*j, &children); | |
5484 | if (num_children == 0) | |
5485 | continue; | |
5486 | for (auto l = children.begin(); l != children.end(); ++l) { | |
5487 | if (*l >= 0) { | |
5488 | ++num; | |
5489 | } else if (num_osds_subtree[*l] > 0) { | |
5490 | num = num + num_osds_subtree[*l]; | |
5491 | } | |
5492 | } | |
5493 | num_osds_subtree[*j] = num; | |
5494 | } | |
5495 | } | |
5496 | num_down_in_osds = down_in_osds.size(); | |
11fdf7f2 | 5497 | ceph_assert(num_down_in_osds <= num_in_osds); |
224ce89b WB |
5498 | if (num_down_in_osds > 0) { |
5499 | // summary of down subtree types and osds | |
5500 | for (int type = max_type; type > 0; type--) { | |
5501 | if (!crush->get_type_name(type)) | |
5502 | continue; | |
5503 | if (subtree_type_down[type].size() > 0) { | |
5504 | ostringstream ss; | |
5505 | ss << subtree_type_down[type].size() << " " | |
5506 | << crush->get_type_name(type); | |
5507 | if (subtree_type_down[type].size() > 1) { | |
5508 | ss << "s"; | |
5509 | } | |
5510 | int sum_down_osds = 0; | |
5511 | for (auto j = subtree_type_down[type].begin(); | |
5512 | j != subtree_type_down[type].end(); | |
5513 | ++j) { | |
5514 | sum_down_osds = sum_down_osds + num_osds_subtree[*j]; | |
5515 | } | |
5516 | ss << " (" << sum_down_osds << " osds) down"; | |
5517 | string err = string("OSD_") + | |
5518 | string(crush->get_type_name(type)) + "_DOWN"; | |
5519 | boost::to_upper(err); | |
5520 | auto& d = checks->add(err, HEALTH_WARN, ss.str()); | |
5521 | for (auto j = subtree_type_down[type].rbegin(); | |
5522 | j != subtree_type_down[type].rend(); | |
5523 | ++j) { | |
5524 | ostringstream ss; | |
5525 | ss << crush->get_type_name(type); | |
5526 | ss << " "; | |
5527 | ss << crush->get_item_name(*j); | |
5528 | // at the top level, do not print location | |
5529 | if (type != max_type) { | |
5530 | ss << " ("; | |
5531 | ss << crush->get_full_location_ordered_string(*j); | |
5532 | ss << ")"; | |
5533 | } | |
5534 | int num = num_osds_subtree[*j]; | |
5535 | ss << " (" << num << " osds)"; | |
5536 | ss << " is down"; | |
5537 | d.detail.push_back(ss.str()); | |
5538 | } | |
5539 | } | |
5540 | } | |
5541 | ostringstream ss; | |
5542 | ss << down_in_osds.size() << " osds down"; | |
5543 | auto& d = checks->add("OSD_DOWN", HEALTH_WARN, ss.str()); | |
5544 | for (auto it = down_in_osds.begin(); it != down_in_osds.end(); ++it) { | |
5545 | ostringstream ss; | |
5546 | ss << "osd." << *it << " ("; | |
5547 | ss << crush->get_full_location_ordered_string(*it); | |
5548 | ss << ") is down"; | |
5549 | d.detail.push_back(ss.str()); | |
5550 | } | |
5551 | } | |
5552 | ||
5553 | if (!osds.empty()) { | |
5554 | ostringstream ss; | |
5555 | ss << osds.size() << " osds exist in the crush map but not in the osdmap"; | |
5556 | auto& d = checks->add("OSD_ORPHAN", HEALTH_WARN, ss.str()); | |
5557 | for (auto osd : osds) { | |
5558 | ostringstream ss; | |
5559 | ss << "osd." << osd << " exists in crush map but not in osdmap"; | |
5560 | d.detail.push_back(ss.str()); | |
5561 | } | |
5562 | } | |
5563 | } | |
5564 | ||
eafe8130 TL |
5565 | std::list<std::string> scrub_messages; |
5566 | bool noscrub = false, nodeepscrub = false; | |
5567 | for (const auto &p : pools) { | |
5568 | if (p.second.flags & pg_pool_t::FLAG_NOSCRUB) { | |
5569 | ostringstream ss; | |
5570 | ss << "Pool " << get_pool_name(p.first) << " has noscrub flag"; | |
5571 | scrub_messages.push_back(ss.str()); | |
5572 | noscrub = true; | |
5573 | } | |
5574 | if (p.second.flags & pg_pool_t::FLAG_NODEEP_SCRUB) { | |
5575 | ostringstream ss; | |
5576 | ss << "Pool " << get_pool_name(p.first) << " has nodeep-scrub flag"; | |
5577 | scrub_messages.push_back(ss.str()); | |
5578 | nodeepscrub = true; | |
5579 | } | |
5580 | } | |
5581 | if (noscrub || nodeepscrub) { | |
5582 | string out = ""; | |
5583 | out += noscrub ? string("noscrub") + (nodeepscrub ? ", " : "") : ""; | |
5584 | out += nodeepscrub ? "nodeep-scrub" : ""; | |
5585 | auto& d = checks->add("POOL_SCRUB_FLAGS", HEALTH_OK, | |
5586 | "Some pool(s) have the " + out + " flag(s) set"); | |
5587 | d.detail.splice(d.detail.end(), scrub_messages); | |
5588 | } | |
5589 | ||
224ce89b WB |
5590 | // OSD_OUT_OF_ORDER_FULL |
5591 | { | |
5592 | // An osd could configure failsafe ratio, to something different | |
5593 | // but for now assume it is the same here. | |
92f5a8d4 | 5594 | float fsr = cct->_conf->osd_failsafe_full_ratio; |
224ce89b WB |
5595 | if (fsr > 1.0) fsr /= 100; |
5596 | float fr = get_full_ratio(); | |
5597 | float br = get_backfillfull_ratio(); | |
5598 | float nr = get_nearfull_ratio(); | |
5599 | ||
5600 | list<string> detail; | |
5601 | // These checks correspond to how OSDService::check_full_status() in an OSD | |
5602 | // handles the improper setting of these values. | |
5603 | if (br < nr) { | |
5604 | ostringstream ss; | |
5605 | ss << "backfillfull_ratio (" << br | |
5606 | << ") < nearfull_ratio (" << nr << "), increased"; | |
5607 | detail.push_back(ss.str()); | |
5608 | br = nr; | |
5609 | } | |
5610 | if (fr < br) { | |
5611 | ostringstream ss; | |
5612 | ss << "full_ratio (" << fr << ") < backfillfull_ratio (" << br | |
5613 | << "), increased"; | |
5614 | detail.push_back(ss.str()); | |
5615 | fr = br; | |
5616 | } | |
5617 | if (fsr < fr) { | |
5618 | ostringstream ss; | |
5619 | ss << "osd_failsafe_full_ratio (" << fsr << ") < full_ratio (" << fr | |
5620 | << "), increased"; | |
5621 | detail.push_back(ss.str()); | |
5622 | } | |
5623 | if (!detail.empty()) { | |
5624 | auto& d = checks->add("OSD_OUT_OF_ORDER_FULL", HEALTH_ERR, | |
5625 | "full ratio(s) out of order"); | |
5626 | d.detail.swap(detail); | |
5627 | } | |
5628 | } | |
5629 | ||
5630 | // OSD_FULL | |
5631 | // OSD_NEARFULL | |
5632 | // OSD_BACKFILLFULL | |
5633 | // OSD_FAILSAFE_FULL | |
5634 | { | |
5635 | set<int> full, backfillfull, nearfull; | |
5636 | get_full_osd_counts(&full, &backfillfull, &nearfull); | |
5637 | if (full.size()) { | |
5638 | ostringstream ss; | |
5639 | ss << full.size() << " full osd(s)"; | |
5640 | auto& d = checks->add("OSD_FULL", HEALTH_ERR, ss.str()); | |
5641 | for (auto& i: full) { | |
5642 | ostringstream ss; | |
5643 | ss << "osd." << i << " is full"; | |
5644 | d.detail.push_back(ss.str()); | |
5645 | } | |
5646 | } | |
5647 | if (backfillfull.size()) { | |
5648 | ostringstream ss; | |
5649 | ss << backfillfull.size() << " backfillfull osd(s)"; | |
5650 | auto& d = checks->add("OSD_BACKFILLFULL", HEALTH_WARN, ss.str()); | |
5651 | for (auto& i: backfillfull) { | |
5652 | ostringstream ss; | |
5653 | ss << "osd." << i << " is backfill full"; | |
5654 | d.detail.push_back(ss.str()); | |
5655 | } | |
5656 | } | |
5657 | if (nearfull.size()) { | |
5658 | ostringstream ss; | |
5659 | ss << nearfull.size() << " nearfull osd(s)"; | |
5660 | auto& d = checks->add("OSD_NEARFULL", HEALTH_WARN, ss.str()); | |
5661 | for (auto& i: nearfull) { | |
5662 | ostringstream ss; | |
5663 | ss << "osd." << i << " is near full"; | |
5664 | d.detail.push_back(ss.str()); | |
5665 | } | |
5666 | } | |
5667 | } | |
5668 | ||
5669 | // OSDMAP_FLAGS | |
5670 | { | |
5671 | // warn about flags | |
5672 | uint64_t warn_flags = | |
3efd9988 | 5673 | CEPH_OSDMAP_NEARFULL | |
224ce89b WB |
5674 | CEPH_OSDMAP_FULL | |
5675 | CEPH_OSDMAP_PAUSERD | | |
5676 | CEPH_OSDMAP_PAUSEWR | | |
5677 | CEPH_OSDMAP_PAUSEREC | | |
5678 | CEPH_OSDMAP_NOUP | | |
5679 | CEPH_OSDMAP_NODOWN | | |
5680 | CEPH_OSDMAP_NOIN | | |
5681 | CEPH_OSDMAP_NOOUT | | |
5682 | CEPH_OSDMAP_NOBACKFILL | | |
5683 | CEPH_OSDMAP_NORECOVER | | |
5684 | CEPH_OSDMAP_NOSCRUB | | |
5685 | CEPH_OSDMAP_NODEEP_SCRUB | | |
5686 | CEPH_OSDMAP_NOTIERAGENT | | |
11fdf7f2 | 5687 | CEPH_OSDMAP_NOSNAPTRIM | |
224ce89b WB |
5688 | CEPH_OSDMAP_NOREBALANCE; |
5689 | if (test_flag(warn_flags)) { | |
5690 | ostringstream ss; | |
5691 | ss << get_flag_string(get_flags() & warn_flags) | |
5692 | << " flag(s) set"; | |
5693 | checks->add("OSDMAP_FLAGS", HEALTH_WARN, ss.str()); | |
5694 | } | |
5695 | } | |
5696 | ||
5697 | // OSD_FLAGS | |
5698 | { | |
5699 | list<string> detail; | |
5700 | const unsigned flags = | |
5701 | CEPH_OSD_NOUP | | |
5702 | CEPH_OSD_NOIN | | |
5703 | CEPH_OSD_NODOWN | | |
5704 | CEPH_OSD_NOOUT; | |
5705 | for (int i = 0; i < max_osd; ++i) { | |
5706 | if (osd_state[i] & flags) { | |
5707 | ostringstream ss; | |
5708 | set<string> states; | |
5709 | OSDMap::calc_state_set(osd_state[i] & flags, states); | |
5710 | ss << "osd." << i << " has flags " << states; | |
5711 | detail.push_back(ss.str()); | |
5712 | } | |
5713 | } | |
81eedcae TL |
5714 | for (auto& i : crush_node_flags) { |
5715 | if (i.second && crush->item_exists(i.first)) { | |
5716 | ostringstream ss; | |
5717 | set<string> states; | |
5718 | OSDMap::calc_state_set(i.second, states); | |
5719 | int t = i.first >= 0 ? 0 : crush->get_bucket_type(i.first); | |
5720 | const char *tn = crush->get_type_name(t); | |
5721 | ss << (tn ? tn : "node") << " " | |
5722 | << crush->get_item_name(i.first) << " has flags " << states; | |
5723 | detail.push_back(ss.str()); | |
5724 | } | |
5725 | } | |
5726 | for (auto& i : device_class_flags) { | |
5727 | const char* class_name = crush->get_class_name(i.first); | |
5728 | if (i.second && class_name) { | |
5729 | ostringstream ss; | |
5730 | set<string> states; | |
5731 | OSDMap::calc_state_set(i.second, states); | |
5732 | ss << "device class '" << class_name << "' has flags " << states; | |
5733 | detail.push_back(ss.str()); | |
5734 | } | |
5735 | } | |
224ce89b WB |
5736 | if (!detail.empty()) { |
5737 | ostringstream ss; | |
81eedcae | 5738 | ss << detail.size() << " OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set"; |
224ce89b WB |
5739 | auto& d = checks->add("OSD_FLAGS", HEALTH_WARN, ss.str()); |
5740 | d.detail.swap(detail); | |
5741 | } | |
5742 | } | |
5743 | ||
5744 | // OLD_CRUSH_TUNABLES | |
92f5a8d4 | 5745 | if (cct->_conf->mon_warn_on_legacy_crush_tunables) { |
224ce89b | 5746 | string min = crush->get_min_required_version(); |
92f5a8d4 | 5747 | if (min < cct->_conf->mon_crush_min_required_version) { |
224ce89b WB |
5748 | ostringstream ss; |
5749 | ss << "crush map has legacy tunables (require " << min | |
92f5a8d4 | 5750 | << ", min is " << cct->_conf->mon_crush_min_required_version << ")"; |
224ce89b WB |
5751 | auto& d = checks->add("OLD_CRUSH_TUNABLES", HEALTH_WARN, ss.str()); |
5752 | d.detail.push_back("see http://docs.ceph.com/docs/master/rados/operations/crush-map/#tunables"); | |
5753 | } | |
5754 | } | |
5755 | ||
5756 | // OLD_CRUSH_STRAW_CALC_VERSION | |
92f5a8d4 | 5757 | if (cct->_conf->mon_warn_on_crush_straw_calc_version_zero) { |
224ce89b WB |
5758 | if (crush->get_straw_calc_version() == 0) { |
5759 | ostringstream ss; | |
5760 | ss << "crush map has straw_calc_version=0"; | |
5761 | auto& d = checks->add("OLD_CRUSH_STRAW_CALC_VERSION", HEALTH_WARN, ss.str()); | |
5762 | d.detail.push_back( | |
5763 | "see http://docs.ceph.com/docs/master/rados/operations/crush-map/#tunables"); | |
5764 | } | |
5765 | } | |
5766 | ||
5767 | // CACHE_POOL_NO_HIT_SET | |
92f5a8d4 | 5768 | if (cct->_conf->mon_warn_on_cache_pools_without_hit_sets) { |
224ce89b WB |
5769 | list<string> detail; |
5770 | for (map<int64_t, pg_pool_t>::const_iterator p = pools.begin(); | |
5771 | p != pools.end(); | |
5772 | ++p) { | |
5773 | const pg_pool_t& info = p->second; | |
5774 | if (info.cache_mode_requires_hit_set() && | |
5775 | info.hit_set_params.get_type() == HitSet::TYPE_NONE) { | |
5776 | ostringstream ss; | |
5777 | ss << "pool '" << get_pool_name(p->first) | |
5778 | << "' with cache_mode " << info.get_cache_mode_name() | |
5779 | << " needs hit_set_type to be set but it is not"; | |
5780 | detail.push_back(ss.str()); | |
5781 | } | |
5782 | } | |
5783 | if (!detail.empty()) { | |
5784 | ostringstream ss; | |
5785 | ss << detail.size() << " cache pools are missing hit_sets"; | |
5786 | auto& d = checks->add("CACHE_POOL_NO_HIT_SET", HEALTH_WARN, ss.str()); | |
5787 | d.detail.swap(detail); | |
5788 | } | |
5789 | } | |
5790 | ||
5791 | // OSD_NO_SORTBITWISE | |
11fdf7f2 | 5792 | if (!test_flag(CEPH_OSDMAP_SORTBITWISE)) { |
224ce89b | 5793 | ostringstream ss; |
11fdf7f2 | 5794 | ss << "'sortbitwise' flag is not set"; |
224ce89b WB |
5795 | checks->add("OSD_NO_SORTBITWISE", HEALTH_WARN, ss.str()); |
5796 | } | |
5797 | ||
5798 | // OSD_UPGRADE_FINISHED | |
5799 | // none of these (yet) since we don't run until luminous upgrade is done. | |
5800 | ||
3efd9988 | 5801 | // POOL_NEARFULL/BACKFILLFULL/FULL |
224ce89b | 5802 | { |
3efd9988 | 5803 | list<string> full_detail, backfillfull_detail, nearfull_detail; |
224ce89b WB |
5804 | for (auto it : get_pools()) { |
5805 | const pg_pool_t &pool = it.second; | |
3efd9988 | 5806 | const string& pool_name = get_pool_name(it.first); |
224ce89b | 5807 | if (pool.has_flag(pg_pool_t::FLAG_FULL)) { |
224ce89b | 5808 | stringstream ss; |
11fdf7f2 | 5809 | if (pool.has_flag(pg_pool_t::FLAG_FULL_QUOTA)) { |
3efd9988 FG |
5810 | // may run out of space too, |
5811 | // but we want EQUOTA taking precedence | |
11fdf7f2 | 5812 | ss << "pool '" << pool_name << "' is full (running out of quota)"; |
3efd9988 FG |
5813 | } else { |
5814 | ss << "pool '" << pool_name << "' is full (no space)"; | |
5815 | } | |
5816 | full_detail.push_back(ss.str()); | |
5817 | } else if (pool.has_flag(pg_pool_t::FLAG_BACKFILLFULL)) { | |
5818 | stringstream ss; | |
5819 | ss << "pool '" << pool_name << "' is backfillfull"; | |
5820 | backfillfull_detail.push_back(ss.str()); | |
5821 | } else if (pool.has_flag(pg_pool_t::FLAG_NEARFULL)) { | |
5822 | stringstream ss; | |
5823 | ss << "pool '" << pool_name << "' is nearfull"; | |
5824 | nearfull_detail.push_back(ss.str()); | |
224ce89b WB |
5825 | } |
5826 | } | |
3efd9988 | 5827 | if (!full_detail.empty()) { |
224ce89b | 5828 | ostringstream ss; |
3efd9988 | 5829 | ss << full_detail.size() << " pool(s) full"; |
224ce89b | 5830 | auto& d = checks->add("POOL_FULL", HEALTH_WARN, ss.str()); |
3efd9988 FG |
5831 | d.detail.swap(full_detail); |
5832 | } | |
5833 | if (!backfillfull_detail.empty()) { | |
5834 | ostringstream ss; | |
5835 | ss << backfillfull_detail.size() << " pool(s) backfillfull"; | |
5836 | auto& d = checks->add("POOL_BACKFILLFULL", HEALTH_WARN, ss.str()); | |
5837 | d.detail.swap(backfillfull_detail); | |
5838 | } | |
5839 | if (!nearfull_detail.empty()) { | |
5840 | ostringstream ss; | |
5841 | ss << nearfull_detail.size() << " pool(s) nearfull"; | |
5842 | auto& d = checks->add("POOL_NEARFULL", HEALTH_WARN, ss.str()); | |
5843 | d.detail.swap(nearfull_detail); | |
224ce89b WB |
5844 | } |
5845 | } | |
92f5a8d4 TL |
5846 | |
5847 | // POOL_PG_NUM_NOT_POWER_OF_TWO | |
5848 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_pg_num_not_power_of_two")) { | |
5849 | list<string> detail; | |
5850 | for (auto it : get_pools()) { | |
5851 | if (!isp2(it.second.get_pg_num_target())) { | |
5852 | ostringstream ss; | |
5853 | ss << "pool '" << get_pool_name(it.first) | |
5854 | << "' pg_num " << it.second.get_pg_num_target() | |
5855 | << " is not a power of two"; | |
5856 | detail.push_back(ss.str()); | |
5857 | } | |
5858 | } | |
5859 | if (!detail.empty()) { | |
5860 | ostringstream ss; | |
5861 | ss << detail.size() << " pool(s) have non-power-of-two pg_num"; | |
5862 | auto& d = checks->add("POOL_PG_NUM_NOT_POWER_OF_TWO", HEALTH_WARN, | |
5863 | ss.str()); | |
5864 | d.detail.swap(detail); | |
5865 | } | |
5866 | } | |
224ce89b | 5867 | } |
35e4c445 FG |
5868 | |
5869 | int OSDMap::parse_osd_id_list(const vector<string>& ls, set<int> *out, | |
5870 | ostream *ss) const | |
5871 | { | |
5872 | out->clear(); | |
5873 | for (auto i = ls.begin(); i != ls.end(); ++i) { | |
5874 | if (i == ls.begin() && | |
5875 | (*i == "any" || *i == "all" || *i == "*")) { | |
5876 | get_all_osds(*out); | |
5877 | break; | |
5878 | } | |
5879 | long osd = parse_osd_id(i->c_str(), ss); | |
5880 | if (osd < 0) { | |
5881 | *ss << "invalid osd id '" << *i << "'"; | |
5882 | return -EINVAL; | |
5883 | } | |
5884 | out->insert(osd); | |
5885 | } | |
5886 | return 0; | |
5887 | } | |
11fdf7f2 TL |
5888 | |
5889 | void OSDMap::get_random_up_osds_by_subtree(int n, // whoami | |
5890 | string &subtree, | |
5891 | int limit, // how many | |
5892 | set<int> skip, | |
5893 | set<int> *want) const { | |
5894 | if (limit <= 0) | |
5895 | return; | |
5896 | int subtree_type = crush->get_type_id(subtree); | |
5897 | if (subtree_type < 1) | |
5898 | return; | |
5899 | vector<int> subtrees; | |
5900 | crush->get_subtree_of_type(subtree_type, &subtrees); | |
5901 | std::random_device rd; | |
5902 | std::default_random_engine rng{rd()}; | |
5903 | std::shuffle(subtrees.begin(), subtrees.end(), rng); | |
5904 | for (auto s : subtrees) { | |
5905 | if (limit <= 0) | |
5906 | break; | |
5907 | if (crush->subtree_contains(s, n)) | |
5908 | continue; | |
5909 | vector<int> osds; | |
5910 | crush->get_children_of_type(s, 0, &osds); | |
5911 | if (osds.empty()) | |
5912 | continue; | |
5913 | vector<int> up_osds; | |
5914 | for (auto o : osds) { | |
5915 | if (is_up(o) && !skip.count(o)) | |
5916 | up_osds.push_back(o); | |
5917 | } | |
5918 | if (up_osds.empty()) | |
5919 | continue; | |
5920 | auto it = up_osds.begin(); | |
5921 | std::advance(it, (n % up_osds.size())); | |
5922 | want->insert(*it); | |
5923 | --limit; | |
5924 | } | |
5925 | } | |
5926 | ||
5927 | float OSDMap::pool_raw_used_rate(int64_t poolid) const | |
5928 | { | |
5929 | const pg_pool_t *pool = get_pg_pool(poolid); | |
5930 | assert(pool != nullptr); | |
5931 | ||
5932 | switch (pool->get_type()) { | |
5933 | case pg_pool_t::TYPE_REPLICATED: | |
5934 | return pool->get_size(); | |
5935 | break; | |
5936 | case pg_pool_t::TYPE_ERASURE: | |
5937 | { | |
5938 | auto& ecp = | |
5939 | get_erasure_code_profile(pool->erasure_code_profile); | |
5940 | auto pm = ecp.find("m"); | |
5941 | auto pk = ecp.find("k"); | |
5942 | if (pm != ecp.end() && pk != ecp.end()) { | |
5943 | int k = atoi(pk->second.c_str()); | |
5944 | int m = atoi(pm->second.c_str()); | |
5945 | int mk = m + k; | |
5946 | ceph_assert(mk != 0); | |
5947 | ceph_assert(k != 0); | |
5948 | return (float)mk / k; | |
5949 | } else { | |
5950 | return 0.0; | |
5951 | } | |
5952 | } | |
5953 | break; | |
5954 | default: | |
5955 | ceph_abort_msg("unrecognized pool type"); | |
5956 | } | |
5957 | } | |
81eedcae TL |
5958 | |
5959 | unsigned OSDMap::get_osd_crush_node_flags(int osd) const | |
5960 | { | |
5961 | unsigned flags = 0; | |
5962 | if (!crush_node_flags.empty()) { | |
5963 | // the map will contain type -> name | |
5964 | std::map<std::string,std::string> ploc = crush->get_full_location(osd); | |
5965 | for (auto& i : ploc) { | |
5966 | int id = crush->get_item_id(i.second); | |
5967 | auto p = crush_node_flags.find(id); | |
5968 | if (p != crush_node_flags.end()) { | |
5969 | flags |= p->second; | |
5970 | } | |
5971 | } | |
5972 | } | |
5973 | return flags; | |
5974 | } | |
5975 | ||
5976 | unsigned OSDMap::get_crush_node_flags(int id) const | |
5977 | { | |
5978 | unsigned flags = 0; | |
5979 | auto it = crush_node_flags.find(id); | |
5980 | if (it != crush_node_flags.end()) | |
5981 | flags = it->second; | |
5982 | return flags; | |
5983 | } | |
5984 | ||
5985 | unsigned OSDMap::get_device_class_flags(int id) const | |
5986 | { | |
5987 | unsigned flags = 0; | |
5988 | auto it = device_class_flags.find(id); | |
5989 | if (it != device_class_flags.end()) | |
5990 | flags = it->second; | |
5991 | return flags; | |
5992 | } |