]>
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" | |
9f95a23c | 30 | #include "include/common_fwd.h" |
7c673cae FG |
31 | #include "include/str_map.h" |
32 | ||
33 | #include "common/code_environment.h" | |
224ce89b | 34 | #include "mon/health_check.h" |
7c673cae FG |
35 | |
36 | #include "crush/CrushTreeDumper.h" | |
37 | #include "common/Clock.h" | |
11fdf7f2 TL |
38 | #include "mon/PGMap.h" |
39 | ||
9f95a23c TL |
40 | using std::list; |
41 | using std::make_pair; | |
42 | using std::map; | |
43 | using std::multimap; | |
44 | using std::ostream; | |
45 | using std::ostringstream; | |
46 | using std::pair; | |
47 | using std::set; | |
48 | using std::string; | |
49 | using std::stringstream; | |
50 | using std::unordered_map; | |
51 | using std::vector; | |
52 | ||
53 | using ceph::decode; | |
54 | using ceph::encode; | |
55 | using ceph::Formatter; | |
56 | ||
7c673cae FG |
57 | #define dout_subsys ceph_subsys_osd |
58 | ||
59 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap, osdmap, osdmap); | |
60 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap::Incremental, osdmap_inc, osdmap); | |
61 | ||
62 | ||
63 | // ---------------------------------- | |
64 | // osd_info_t | |
65 | ||
66 | void osd_info_t::dump(Formatter *f) const | |
67 | { | |
68 | f->dump_int("last_clean_begin", last_clean_begin); | |
69 | f->dump_int("last_clean_end", last_clean_end); | |
70 | f->dump_int("up_from", up_from); | |
71 | f->dump_int("up_thru", up_thru); | |
72 | f->dump_int("down_at", down_at); | |
73 | f->dump_int("lost_at", lost_at); | |
74 | } | |
75 | ||
9f95a23c | 76 | void osd_info_t::encode(ceph::buffer::list& bl) const |
7c673cae | 77 | { |
11fdf7f2 | 78 | using ceph::encode; |
7c673cae | 79 | __u8 struct_v = 1; |
11fdf7f2 TL |
80 | encode(struct_v, bl); |
81 | encode(last_clean_begin, bl); | |
82 | encode(last_clean_end, bl); | |
83 | encode(up_from, bl); | |
84 | encode(up_thru, bl); | |
85 | encode(down_at, bl); | |
86 | encode(lost_at, bl); | |
7c673cae FG |
87 | } |
88 | ||
9f95a23c | 89 | void osd_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 90 | { |
11fdf7f2 | 91 | using ceph::decode; |
7c673cae | 92 | __u8 struct_v; |
11fdf7f2 TL |
93 | decode(struct_v, bl); |
94 | decode(last_clean_begin, bl); | |
95 | decode(last_clean_end, bl); | |
96 | decode(up_from, bl); | |
97 | decode(up_thru, bl); | |
98 | decode(down_at, bl); | |
99 | decode(lost_at, bl); | |
7c673cae FG |
100 | } |
101 | ||
102 | void osd_info_t::generate_test_instances(list<osd_info_t*>& o) | |
103 | { | |
104 | o.push_back(new osd_info_t); | |
105 | o.push_back(new osd_info_t); | |
106 | o.back()->last_clean_begin = 1; | |
107 | o.back()->last_clean_end = 2; | |
108 | o.back()->up_from = 30; | |
109 | o.back()->up_thru = 40; | |
110 | o.back()->down_at = 5; | |
111 | o.back()->lost_at = 6; | |
112 | } | |
113 | ||
114 | ostream& operator<<(ostream& out, const osd_info_t& info) | |
115 | { | |
116 | out << "up_from " << info.up_from | |
117 | << " up_thru " << info.up_thru | |
118 | << " down_at " << info.down_at | |
119 | << " last_clean_interval [" << info.last_clean_begin << "," << info.last_clean_end << ")"; | |
120 | if (info.lost_at) | |
121 | out << " lost_at " << info.lost_at; | |
122 | return out; | |
123 | } | |
124 | ||
125 | // ---------------------------------- | |
126 | // osd_xinfo_t | |
127 | ||
128 | void osd_xinfo_t::dump(Formatter *f) const | |
129 | { | |
130 | f->dump_stream("down_stamp") << down_stamp; | |
131 | f->dump_float("laggy_probability", laggy_probability); | |
132 | f->dump_int("laggy_interval", laggy_interval); | |
133 | f->dump_int("features", features); | |
134 | f->dump_unsigned("old_weight", old_weight); | |
9f95a23c TL |
135 | f->dump_stream("last_purged_snaps_scrub") << last_purged_snaps_scrub; |
136 | f->dump_int("dead_epoch", dead_epoch); | |
7c673cae FG |
137 | } |
138 | ||
9f95a23c | 139 | void osd_xinfo_t::encode(ceph::buffer::list& bl, uint64_t enc_features) const |
7c673cae | 140 | { |
9f95a23c TL |
141 | uint8_t v = 4; |
142 | if (!HAVE_FEATURE(enc_features, SERVER_OCTOPUS)) { | |
143 | v = 3; | |
144 | } | |
145 | ENCODE_START(v, 1, bl); | |
11fdf7f2 | 146 | encode(down_stamp, bl); |
f67539c2 | 147 | __u32 lp = laggy_probability * float(0xfffffffful); |
11fdf7f2 TL |
148 | encode(lp, bl); |
149 | encode(laggy_interval, bl); | |
150 | encode(features, bl); | |
151 | encode(old_weight, bl); | |
9f95a23c TL |
152 | if (v >= 4) { |
153 | encode(last_purged_snaps_scrub, bl); | |
154 | encode(dead_epoch, bl); | |
155 | } | |
7c673cae FG |
156 | ENCODE_FINISH(bl); |
157 | } | |
158 | ||
9f95a23c | 159 | void osd_xinfo_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 160 | { |
9f95a23c | 161 | DECODE_START(4, bl); |
11fdf7f2 | 162 | decode(down_stamp, bl); |
7c673cae | 163 | __u32 lp; |
11fdf7f2 | 164 | decode(lp, bl); |
7c673cae | 165 | laggy_probability = (float)lp / (float)0xffffffff; |
11fdf7f2 | 166 | decode(laggy_interval, bl); |
7c673cae | 167 | if (struct_v >= 2) |
11fdf7f2 | 168 | decode(features, bl); |
7c673cae FG |
169 | else |
170 | features = 0; | |
171 | if (struct_v >= 3) | |
11fdf7f2 | 172 | decode(old_weight, bl); |
7c673cae FG |
173 | else |
174 | old_weight = 0; | |
9f95a23c TL |
175 | if (struct_v >= 4) { |
176 | decode(last_purged_snaps_scrub, bl); | |
177 | decode(dead_epoch, bl); | |
178 | } else { | |
179 | dead_epoch = 0; | |
180 | } | |
7c673cae FG |
181 | DECODE_FINISH(bl); |
182 | } | |
183 | ||
184 | void osd_xinfo_t::generate_test_instances(list<osd_xinfo_t*>& o) | |
185 | { | |
186 | o.push_back(new osd_xinfo_t); | |
187 | o.push_back(new osd_xinfo_t); | |
188 | o.back()->down_stamp = utime_t(2, 3); | |
189 | o.back()->laggy_probability = .123; | |
190 | o.back()->laggy_interval = 123456; | |
191 | o.back()->old_weight = 0x7fff; | |
192 | } | |
193 | ||
194 | ostream& operator<<(ostream& out, const osd_xinfo_t& xi) | |
195 | { | |
196 | return out << "down_stamp " << xi.down_stamp | |
197 | << " laggy_probability " << xi.laggy_probability | |
198 | << " laggy_interval " << xi.laggy_interval | |
9f95a23c TL |
199 | << " old_weight " << xi.old_weight |
200 | << " last_purged_snaps_scrub " << xi.last_purged_snaps_scrub | |
201 | << " dead_epoch " << xi.dead_epoch; | |
7c673cae FG |
202 | } |
203 | ||
204 | // ---------------------------------- | |
205 | // OSDMap::Incremental | |
206 | ||
207 | int OSDMap::Incremental::get_net_marked_out(const OSDMap *previous) const | |
208 | { | |
209 | int n = 0; | |
210 | for (auto &weight : new_weight) { | |
211 | if (weight.second == CEPH_OSD_OUT && !previous->is_out(weight.first)) | |
212 | n++; // marked out | |
213 | else if (weight.second != CEPH_OSD_OUT && previous->is_out(weight.first)) | |
214 | n--; // marked in | |
215 | } | |
216 | return n; | |
217 | } | |
218 | ||
219 | int OSDMap::Incremental::get_net_marked_down(const OSDMap *previous) const | |
220 | { | |
221 | int n = 0; | |
222 | for (auto &state : new_state) { // | |
223 | if (state.second & CEPH_OSD_UP) { | |
224 | if (previous->is_up(state.first)) | |
225 | n++; // marked down | |
226 | else | |
227 | n--; // marked up | |
228 | } | |
229 | } | |
230 | return n; | |
231 | } | |
232 | ||
233 | int OSDMap::Incremental::identify_osd(uuid_d u) const | |
234 | { | |
235 | for (auto &uuid : new_uuid) | |
236 | if (uuid.second == u) | |
237 | return uuid.first; | |
238 | return -1; | |
239 | } | |
240 | ||
f67539c2 TL |
241 | int OSDMap::Incremental::propagate_base_properties_to_tiers(CephContext *cct, |
242 | const OSDMap& osdmap) | |
7c673cae | 243 | { |
11fdf7f2 | 244 | ceph_assert(epoch == osdmap.get_epoch() + 1); |
7c673cae FG |
245 | |
246 | for (auto &new_pool : new_pools) { | |
247 | if (!new_pool.second.tiers.empty()) { | |
248 | pg_pool_t& base = new_pool.second; | |
249 | ||
11fdf7f2 TL |
250 | auto new_rem_it = new_removed_snaps.find(new_pool.first); |
251 | ||
7c673cae FG |
252 | for (const auto &tier_pool : base.tiers) { |
253 | const auto &r = new_pools.find(tier_pool); | |
254 | pg_pool_t *tier = 0; | |
255 | if (r == new_pools.end()) { | |
256 | const pg_pool_t *orig = osdmap.get_pg_pool(tier_pool); | |
257 | if (!orig) { | |
258 | lderr(cct) << __func__ << " no pool " << tier_pool << dendl; | |
259 | return -EIO; | |
260 | } | |
261 | tier = get_new_pool(tier_pool, orig); | |
262 | } else { | |
263 | tier = &r->second; | |
264 | } | |
265 | if (tier->tier_of != new_pool.first) { | |
266 | lderr(cct) << __func__ << " " << r->first << " tier_of != " << new_pool.first << dendl; | |
267 | return -EIO; | |
268 | } | |
269 | ||
270 | ldout(cct, 10) << __func__ << " from " << new_pool.first << " to " | |
271 | << tier_pool << dendl; | |
272 | tier->snap_seq = base.snap_seq; | |
273 | tier->snap_epoch = base.snap_epoch; | |
274 | tier->snaps = base.snaps; | |
275 | tier->removed_snaps = base.removed_snaps; | |
11fdf7f2 TL |
276 | tier->flags |= base.flags & (pg_pool_t::FLAG_SELFMANAGED_SNAPS| |
277 | pg_pool_t::FLAG_POOL_SNAPS); | |
278 | ||
279 | if (new_rem_it != new_removed_snaps.end()) { | |
280 | new_removed_snaps[tier_pool] = new_rem_it->second; | |
281 | } | |
f67539c2 TL |
282 | |
283 | tier->application_metadata = base.application_metadata; | |
7c673cae FG |
284 | } |
285 | } | |
286 | } | |
287 | return 0; | |
288 | } | |
289 | ||
28e407b8 AA |
290 | // ---------------------------------- |
291 | // OSDMap | |
7c673cae FG |
292 | |
293 | bool OSDMap::subtree_is_down(int id, set<int> *down_cache) const | |
294 | { | |
295 | if (id >= 0) | |
296 | return is_down(id); | |
297 | ||
298 | if (down_cache && | |
299 | down_cache->count(id)) { | |
300 | return true; | |
301 | } | |
302 | ||
303 | list<int> children; | |
304 | crush->get_children(id, &children); | |
305 | for (const auto &child : children) { | |
306 | if (!subtree_is_down(child, down_cache)) { | |
307 | return false; | |
308 | } | |
309 | } | |
310 | if (down_cache) { | |
311 | down_cache->insert(id); | |
312 | } | |
313 | return true; | |
314 | } | |
315 | ||
316 | bool OSDMap::containing_subtree_is_down(CephContext *cct, int id, int subtree_type, set<int> *down_cache) const | |
317 | { | |
318 | // use a stack-local down_cache if we didn't get one from the | |
319 | // caller. then at least this particular call will avoid duplicated | |
320 | // work. | |
321 | set<int> local_down_cache; | |
322 | if (!down_cache) { | |
323 | down_cache = &local_down_cache; | |
324 | } | |
325 | ||
326 | int current = id; | |
327 | while (true) { | |
328 | int type; | |
329 | if (current >= 0) { | |
330 | type = 0; | |
331 | } else { | |
332 | type = crush->get_bucket_type(current); | |
333 | } | |
11fdf7f2 | 334 | ceph_assert(type >= 0); |
7c673cae FG |
335 | |
336 | if (!subtree_is_down(current, down_cache)) { | |
337 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = false" << dendl; | |
338 | return false; | |
339 | } | |
340 | ||
341 | // is this a big enough subtree to be marked as down? | |
342 | if (type >= subtree_type) { | |
343 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = true ... " << type << " >= " << subtree_type << dendl; | |
344 | return true; | |
345 | } | |
346 | ||
347 | int r = crush->get_immediate_parent_id(current, ¤t); | |
348 | if (r < 0) { | |
349 | return false; | |
350 | } | |
351 | } | |
352 | } | |
353 | ||
224ce89b WB |
354 | bool OSDMap::subtree_type_is_down( |
355 | CephContext *cct, | |
356 | int id, | |
357 | int subtree_type, | |
358 | set<int> *down_in_osds, | |
359 | set<int> *up_in_osds, | |
360 | set<int> *subtree_up, | |
361 | unordered_map<int, set<int> > *subtree_type_down) const | |
31f18b77 FG |
362 | { |
363 | if (id >= 0) { | |
364 | bool is_down_ret = is_down(id); | |
365 | if (!is_out(id)) { | |
366 | if (is_down_ret) { | |
367 | down_in_osds->insert(id); | |
368 | } else { | |
369 | up_in_osds->insert(id); | |
370 | } | |
371 | } | |
372 | return is_down_ret; | |
373 | } | |
374 | ||
375 | if (subtree_type_down && | |
376 | (*subtree_type_down)[subtree_type].count(id)) { | |
377 | return true; | |
378 | } | |
379 | ||
380 | list<int> children; | |
381 | crush->get_children(id, &children); | |
382 | for (const auto &child : children) { | |
224ce89b WB |
383 | if (!subtree_type_is_down( |
384 | cct, child, crush->get_bucket_type(child), | |
385 | down_in_osds, up_in_osds, subtree_up, subtree_type_down)) { | |
31f18b77 FG |
386 | subtree_up->insert(id); |
387 | return false; | |
388 | } | |
389 | } | |
390 | if (subtree_type_down) { | |
391 | (*subtree_type_down)[subtree_type].insert(id); | |
392 | } | |
393 | return true; | |
394 | } | |
395 | ||
9f95a23c | 396 | void OSDMap::Incremental::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 397 | { |
11fdf7f2 | 398 | using ceph::encode; |
7c673cae | 399 | __u16 v = 5; |
11fdf7f2 TL |
400 | encode(v, bl); |
401 | encode(fsid, bl); | |
402 | encode(epoch, bl); | |
403 | encode(modified, bl); | |
7c673cae | 404 | int32_t new_t = new_pool_max; |
11fdf7f2 TL |
405 | encode(new_t, bl); |
406 | encode(new_flags, bl); | |
407 | encode(fullmap, bl); | |
408 | encode(crush, bl); | |
7c673cae | 409 | |
11fdf7f2 TL |
410 | encode(new_max_osd, bl); |
411 | // for encode(new_pools, bl); | |
7c673cae | 412 | __u32 n = new_pools.size(); |
11fdf7f2 | 413 | encode(n, bl); |
7c673cae FG |
414 | for (const auto &new_pool : new_pools) { |
415 | n = new_pool.first; | |
11fdf7f2 TL |
416 | encode(n, bl); |
417 | encode(new_pool.second, bl, 0); | |
7c673cae | 418 | } |
11fdf7f2 | 419 | // for encode(new_pool_names, bl); |
7c673cae | 420 | n = new_pool_names.size(); |
11fdf7f2 | 421 | encode(n, bl); |
7c673cae FG |
422 | |
423 | for (const auto &new_pool_name : new_pool_names) { | |
424 | n = new_pool_name.first; | |
11fdf7f2 TL |
425 | encode(n, bl); |
426 | encode(new_pool_name.second, bl); | |
7c673cae | 427 | } |
11fdf7f2 | 428 | // for encode(old_pools, bl); |
7c673cae | 429 | n = old_pools.size(); |
11fdf7f2 | 430 | encode(n, bl); |
7c673cae FG |
431 | for (auto &old_pool : old_pools) { |
432 | n = old_pool; | |
11fdf7f2 | 433 | encode(n, bl); |
7c673cae | 434 | } |
11fdf7f2 | 435 | encode(new_up_client, bl, 0); |
31f18b77 FG |
436 | { |
437 | // legacy is map<int32_t,uint8_t> | |
9f95a23c | 438 | map<int32_t, uint8_t> os; |
31f18b77 | 439 | for (auto p : new_state) { |
9f95a23c TL |
440 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
441 | // that an old client could not understand. | |
442 | // skip those! | |
443 | uint8_t s = p.second; | |
444 | if (p.second != 0 && s == 0) | |
445 | continue; | |
446 | os[p.first] = s; | |
447 | } | |
448 | uint32_t n = os.size(); | |
449 | encode(n, bl); | |
450 | for (auto p : os) { | |
11fdf7f2 | 451 | encode(p.first, bl); |
9f95a23c | 452 | encode(p.second, bl); |
31f18b77 FG |
453 | } |
454 | } | |
11fdf7f2 TL |
455 | encode(new_weight, bl); |
456 | // for encode(new_pg_temp, bl); | |
7c673cae | 457 | n = new_pg_temp.size(); |
11fdf7f2 | 458 | encode(n, bl); |
7c673cae FG |
459 | |
460 | for (const auto &pg_temp : new_pg_temp) { | |
461 | old_pg_t opg = pg_temp.first.get_old_pg(); | |
11fdf7f2 TL |
462 | encode(opg, bl); |
463 | encode(pg_temp.second, bl); | |
7c673cae FG |
464 | } |
465 | } | |
466 | ||
9f95a23c | 467 | void OSDMap::Incremental::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 468 | { |
11fdf7f2 | 469 | using ceph::encode; |
7c673cae FG |
470 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
471 | encode_client_old(bl); | |
472 | return; | |
473 | } | |
474 | ||
475 | // base | |
476 | __u16 v = 6; | |
11fdf7f2 TL |
477 | encode(v, bl); |
478 | encode(fsid, bl); | |
479 | encode(epoch, bl); | |
480 | encode(modified, bl); | |
481 | encode(new_pool_max, bl); | |
482 | encode(new_flags, bl); | |
483 | encode(fullmap, bl); | |
484 | encode(crush, bl); | |
485 | ||
486 | encode(new_max_osd, bl); | |
487 | encode(new_pools, bl, features); | |
488 | encode(new_pool_names, bl); | |
489 | encode(old_pools, bl); | |
490 | encode(new_up_client, bl, features); | |
31f18b77 | 491 | { |
9f95a23c | 492 | map<int32_t, uint8_t> os; |
31f18b77 | 493 | for (auto p : new_state) { |
9f95a23c TL |
494 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
495 | // that an old client could not understand. | |
496 | // skip those! | |
497 | uint8_t s = p.second; | |
498 | if (p.second != 0 && s == 0) | |
499 | continue; | |
500 | os[p.first] = s; | |
501 | } | |
502 | uint32_t n = os.size(); | |
503 | encode(n, bl); | |
504 | for (auto p : os) { | |
11fdf7f2 | 505 | encode(p.first, bl); |
9f95a23c | 506 | encode(p.second, bl); |
31f18b77 FG |
507 | } |
508 | } | |
11fdf7f2 TL |
509 | encode(new_weight, bl); |
510 | encode(new_pg_temp, bl); | |
7c673cae FG |
511 | |
512 | // extended | |
513 | __u16 ev = 10; | |
11fdf7f2 TL |
514 | encode(ev, bl); |
515 | encode(new_hb_back_up, bl, features); | |
516 | encode(new_up_thru, bl); | |
517 | encode(new_last_clean_interval, bl); | |
518 | encode(new_lost, bl); | |
f67539c2 TL |
519 | encode(new_blocklist, bl, features); |
520 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
521 | encode(new_up_cluster, bl, features); |
522 | encode(cluster_snapshot, bl); | |
523 | encode(new_uuid, bl); | |
9f95a23c | 524 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
525 | encode(new_hb_front_up, bl, features); |
526 | } | |
527 | ||
528 | template<class T> | |
9f95a23c | 529 | static void encode_addrvec_map_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
530 | { |
531 | uint32_t n = m.size(); | |
532 | encode(n, bl); | |
533 | for (auto& i : m) { | |
534 | encode(i.first, bl); | |
535 | encode(i.second.legacy_addr(), bl, f); | |
536 | } | |
537 | } | |
538 | ||
539 | template<class T> | |
9f95a23c | 540 | static void encode_addrvec_pvec_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
541 | { |
542 | uint32_t n = m.size(); | |
543 | encode(n, bl); | |
544 | for (auto& i : m) { | |
545 | if (i) { | |
546 | encode(i->legacy_addr(), bl, f); | |
547 | } else { | |
548 | encode(entity_addr_t(), bl, f); | |
549 | } | |
550 | } | |
7c673cae FG |
551 | } |
552 | ||
11fdf7f2 TL |
553 | /* for a description of osdmap incremental versions, and when they were |
554 | * introduced, please refer to | |
555 | * doc/dev/osd_internals/osdmap_versions.txt | |
556 | */ | |
9f95a23c | 557 | void OSDMap::Incremental::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 558 | { |
11fdf7f2 | 559 | using ceph::encode; |
7c673cae FG |
560 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
561 | encode_classic(bl, features); | |
562 | return; | |
563 | } | |
564 | ||
565 | // only a select set of callers should *ever* be encoding new | |
566 | // OSDMaps. others should be passing around the canonical encoded | |
567 | // buffers from on high. select out those callers by passing in an | |
568 | // "impossible" feature bit. | |
11fdf7f2 | 569 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
570 | features &= ~CEPH_FEATURE_RESERVED; |
571 | ||
572 | size_t start_offset = bl.length(); | |
573 | size_t tail_offset; | |
11fdf7f2 | 574 | size_t crc_offset; |
9f95a23c | 575 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
576 | |
577 | // meta-encoding: how we include client-used and osd-specific data | |
578 | ENCODE_START(8, 7, bl); | |
579 | ||
580 | { | |
11fdf7f2 | 581 | uint8_t v = 8; |
7c673cae FG |
582 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
583 | v = 3; | |
11fdf7f2 TL |
584 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
585 | v = 5; | |
586 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
587 | v = 6; | |
7c673cae FG |
588 | } |
589 | ENCODE_START(v, 1, bl); // client-usable data | |
11fdf7f2 TL |
590 | encode(fsid, bl); |
591 | encode(epoch, bl); | |
592 | encode(modified, bl); | |
593 | encode(new_pool_max, bl); | |
594 | encode(new_flags, bl); | |
595 | encode(fullmap, bl); | |
596 | encode(crush, bl); | |
597 | ||
598 | encode(new_max_osd, bl); | |
599 | encode(new_pools, bl, features); | |
600 | encode(new_pool_names, bl); | |
601 | encode(old_pools, bl); | |
602 | if (v >= 7) { | |
603 | encode(new_up_client, bl, features); | |
604 | } else { | |
605 | encode_addrvec_map_as_addr(new_up_client, bl, features); | |
606 | } | |
31f18b77 | 607 | if (v >= 5) { |
11fdf7f2 | 608 | encode(new_state, bl); |
31f18b77 | 609 | } else { |
9f95a23c | 610 | map<int32_t, uint8_t> os; |
31f18b77 | 611 | for (auto p : new_state) { |
9f95a23c TL |
612 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
613 | // that an old client could not understand. | |
614 | // skip those! | |
615 | uint8_t s = p.second; | |
616 | if (p.second != 0 && s == 0) | |
617 | continue; | |
618 | os[p.first] = s; | |
619 | } | |
620 | uint32_t n = os.size(); | |
621 | encode(n, bl); | |
622 | for (auto p : os) { | |
623 | encode(p.first, bl); | |
624 | encode(p.second, bl); | |
31f18b77 FG |
625 | } |
626 | } | |
11fdf7f2 TL |
627 | encode(new_weight, bl); |
628 | encode(new_pg_temp, bl); | |
629 | encode(new_primary_temp, bl); | |
630 | encode(new_primary_affinity, bl); | |
631 | encode(new_erasure_code_profiles, bl); | |
632 | encode(old_erasure_code_profiles, bl); | |
7c673cae | 633 | if (v >= 4) { |
11fdf7f2 TL |
634 | encode(new_pg_upmap, bl); |
635 | encode(old_pg_upmap, bl); | |
636 | encode(new_pg_upmap_items, bl); | |
637 | encode(old_pg_upmap_items, bl); | |
638 | } | |
639 | if (v >= 6) { | |
640 | encode(new_removed_snaps, bl); | |
641 | encode(new_purged_snaps, bl); | |
642 | } | |
643 | if (v >= 8) { | |
644 | encode(new_last_up_change, bl); | |
645 | encode(new_last_in_change, bl); | |
7c673cae FG |
646 | } |
647 | ENCODE_FINISH(bl); // client-usable data | |
648 | } | |
649 | ||
650 | { | |
f67539c2 | 651 | uint8_t target_v = 9; // if bumping this, be aware of stretch_mode target_v 10! |
7c673cae FG |
652 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
653 | target_v = 2; | |
11fdf7f2 TL |
654 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { |
655 | target_v = 6; | |
7c673cae | 656 | } |
f67539c2 | 657 | if (change_stretch_mode) { |
f67539c2 TL |
658 | target_v = std::max((uint8_t)10, target_v); |
659 | } | |
7c673cae | 660 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
661 | if (target_v < 7) { |
662 | encode_addrvec_map_as_addr(new_hb_back_up, bl, features); | |
663 | } else { | |
664 | encode(new_hb_back_up, bl, features); | |
665 | } | |
666 | encode(new_up_thru, bl); | |
667 | encode(new_last_clean_interval, bl); | |
668 | encode(new_lost, bl); | |
f67539c2 TL |
669 | encode(new_blocklist, bl, features); |
670 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
671 | if (target_v < 7) { |
672 | encode_addrvec_map_as_addr(new_up_cluster, bl, features); | |
673 | } else { | |
674 | encode(new_up_cluster, bl, features); | |
675 | } | |
676 | encode(cluster_snapshot, bl); | |
677 | encode(new_uuid, bl); | |
9f95a23c | 678 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
679 | if (target_v < 7) { |
680 | encode_addrvec_map_as_addr(new_hb_front_up, bl, features); | |
681 | } else { | |
682 | encode(new_hb_front_up, bl, features); | |
683 | } | |
684 | encode(features, bl); // NOTE: features arg, not the member | |
7c673cae | 685 | if (target_v >= 3) { |
11fdf7f2 TL |
686 | encode(new_nearfull_ratio, bl); |
687 | encode(new_full_ratio, bl); | |
688 | encode(new_backfillfull_ratio, bl); | |
31f18b77 FG |
689 | } |
690 | // 5 was string-based new_require_min_compat_client | |
691 | if (target_v >= 6) { | |
11fdf7f2 TL |
692 | encode(new_require_min_compat_client, bl); |
693 | encode(new_require_osd_release, bl); | |
7c673cae | 694 | } |
81eedcae TL |
695 | if (target_v >= 8) { |
696 | encode(new_crush_node_flags, bl); | |
697 | } | |
698 | if (target_v >= 9) { | |
699 | encode(new_device_class_flags, bl); | |
700 | } | |
f67539c2 TL |
701 | if (target_v >= 10) { |
702 | encode(change_stretch_mode, bl); | |
703 | encode(new_stretch_bucket_count, bl); | |
704 | encode(new_degraded_stretch_mode, bl); | |
705 | encode(new_recovering_stretch_mode, bl); | |
706 | encode(new_stretch_mode_bucket, bl); | |
707 | encode(stretch_mode_enabled, bl); | |
708 | } | |
7c673cae FG |
709 | ENCODE_FINISH(bl); // osd-only data |
710 | } | |
711 | ||
11fdf7f2 TL |
712 | crc_offset = bl.length(); |
713 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
714 | tail_offset = bl.length(); |
715 | ||
11fdf7f2 | 716 | encode(full_crc, bl); |
7c673cae FG |
717 | |
718 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
719 | ||
720 | // fill in crc | |
9f95a23c | 721 | ceph::buffer::list front; |
11fdf7f2 | 722 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae | 723 | inc_crc = front.crc32c(-1); |
9f95a23c | 724 | ceph::buffer::list tail; |
7c673cae FG |
725 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
726 | inc_crc = tail.crc32c(inc_crc); | |
727 | ceph_le32 crc_le; | |
728 | crc_le = inc_crc; | |
11fdf7f2 | 729 | crc_filler->copy_in(4u, (char*)&crc_le); |
7c673cae FG |
730 | have_crc = true; |
731 | } | |
732 | ||
9f95a23c | 733 | void OSDMap::Incremental::decode_classic(ceph::buffer::list::const_iterator &p) |
7c673cae | 734 | { |
11fdf7f2 | 735 | using ceph::decode; |
7c673cae FG |
736 | __u32 n, t; |
737 | // base | |
738 | __u16 v; | |
11fdf7f2 TL |
739 | decode(v, p); |
740 | decode(fsid, p); | |
741 | decode(epoch, p); | |
742 | decode(modified, p); | |
7c673cae | 743 | if (v == 4 || v == 5) { |
11fdf7f2 | 744 | decode(n, p); |
7c673cae FG |
745 | new_pool_max = n; |
746 | } else if (v >= 6) | |
11fdf7f2 TL |
747 | decode(new_pool_max, p); |
748 | decode(new_flags, p); | |
749 | decode(fullmap, p); | |
750 | decode(crush, p); | |
7c673cae | 751 | |
11fdf7f2 | 752 | decode(new_max_osd, p); |
7c673cae FG |
753 | if (v < 6) { |
754 | new_pools.clear(); | |
11fdf7f2 | 755 | decode(n, p); |
7c673cae | 756 | while (n--) { |
11fdf7f2 TL |
757 | decode(t, p); |
758 | decode(new_pools[t], p); | |
7c673cae FG |
759 | } |
760 | } else { | |
11fdf7f2 | 761 | decode(new_pools, p); |
7c673cae FG |
762 | } |
763 | if (v == 5) { | |
764 | new_pool_names.clear(); | |
11fdf7f2 | 765 | decode(n, p); |
7c673cae | 766 | while (n--) { |
11fdf7f2 TL |
767 | decode(t, p); |
768 | decode(new_pool_names[t], p); | |
7c673cae FG |
769 | } |
770 | } else if (v >= 6) { | |
11fdf7f2 | 771 | decode(new_pool_names, p); |
7c673cae FG |
772 | } |
773 | if (v < 6) { | |
774 | old_pools.clear(); | |
11fdf7f2 | 775 | decode(n, p); |
7c673cae | 776 | while (n--) { |
11fdf7f2 | 777 | decode(t, p); |
7c673cae FG |
778 | old_pools.insert(t); |
779 | } | |
780 | } else { | |
11fdf7f2 | 781 | decode(old_pools, p); |
7c673cae | 782 | } |
11fdf7f2 | 783 | decode(new_up_client, p); |
31f18b77 FG |
784 | { |
785 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 786 | decode(ns, p); |
31f18b77 FG |
787 | for (auto q : ns) { |
788 | new_state[q.first] = q.second; | |
789 | } | |
790 | } | |
11fdf7f2 | 791 | decode(new_weight, p); |
7c673cae FG |
792 | |
793 | if (v < 6) { | |
794 | new_pg_temp.clear(); | |
11fdf7f2 | 795 | decode(n, p); |
7c673cae FG |
796 | while (n--) { |
797 | old_pg_t opg; | |
9f95a23c | 798 | ceph::decode_raw(opg, p); |
11fdf7f2 | 799 | decode(new_pg_temp[pg_t(opg)], p); |
7c673cae FG |
800 | } |
801 | } else { | |
11fdf7f2 | 802 | decode(new_pg_temp, p); |
7c673cae FG |
803 | } |
804 | ||
805 | // decode short map, too. | |
806 | if (v == 5 && p.end()) | |
807 | return; | |
808 | ||
809 | // extended | |
810 | __u16 ev = 0; | |
811 | if (v >= 5) | |
11fdf7f2 TL |
812 | decode(ev, p); |
813 | decode(new_hb_back_up, p); | |
7c673cae | 814 | if (v < 5) |
11fdf7f2 TL |
815 | decode(new_pool_names, p); |
816 | decode(new_up_thru, p); | |
817 | decode(new_last_clean_interval, p); | |
818 | decode(new_lost, p); | |
f67539c2 TL |
819 | decode(new_blocklist, p); |
820 | decode(old_blocklist, p); | |
7c673cae | 821 | if (ev >= 6) |
11fdf7f2 | 822 | decode(new_up_cluster, p); |
7c673cae | 823 | if (ev >= 7) |
11fdf7f2 | 824 | decode(cluster_snapshot, p); |
7c673cae | 825 | if (ev >= 8) |
11fdf7f2 | 826 | decode(new_uuid, p); |
7c673cae | 827 | if (ev >= 9) |
11fdf7f2 | 828 | decode(new_xinfo, p); |
7c673cae | 829 | if (ev >= 10) |
11fdf7f2 | 830 | decode(new_hb_front_up, p); |
7c673cae FG |
831 | } |
832 | ||
11fdf7f2 TL |
833 | /* for a description of osdmap incremental versions, and when they were |
834 | * introduced, please refer to | |
835 | * doc/dev/osd_internals/osdmap_versions.txt | |
836 | */ | |
9f95a23c | 837 | void OSDMap::Incremental::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 838 | { |
11fdf7f2 | 839 | using ceph::decode; |
7c673cae FG |
840 | /** |
841 | * Older encodings of the Incremental had a single struct_v which | |
842 | * covered the whole encoding, and was prior to our modern | |
843 | * stuff which includes a compatv and a size. So if we see | |
844 | * a struct_v < 7, we must rewind to the beginning and use our | |
845 | * classic decoder. | |
846 | */ | |
847 | size_t start_offset = bl.get_off(); | |
848 | size_t tail_offset = 0; | |
9f95a23c | 849 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
850 | |
851 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
852 | if (struct_v < 7) { | |
11fdf7f2 | 853 | bl.seek(start_offset); |
7c673cae FG |
854 | decode_classic(bl); |
855 | encode_features = 0; | |
856 | if (struct_v >= 6) | |
857 | encode_features = CEPH_FEATURE_PGID64; | |
858 | else | |
859 | encode_features = 0; | |
860 | return; | |
861 | } | |
862 | { | |
11fdf7f2 TL |
863 | DECODE_START(8, bl); // client-usable data |
864 | decode(fsid, bl); | |
865 | decode(epoch, bl); | |
866 | decode(modified, bl); | |
867 | decode(new_pool_max, bl); | |
868 | decode(new_flags, bl); | |
869 | decode(fullmap, bl); | |
870 | decode(crush, bl); | |
871 | ||
872 | decode(new_max_osd, bl); | |
873 | decode(new_pools, bl); | |
874 | decode(new_pool_names, bl); | |
875 | decode(old_pools, bl); | |
876 | decode(new_up_client, bl); | |
31f18b77 | 877 | if (struct_v >= 5) { |
11fdf7f2 | 878 | decode(new_state, bl); |
31f18b77 FG |
879 | } else { |
880 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 881 | decode(ns, bl); |
31f18b77 FG |
882 | for (auto q : ns) { |
883 | new_state[q.first] = q.second; | |
884 | } | |
885 | } | |
11fdf7f2 TL |
886 | decode(new_weight, bl); |
887 | decode(new_pg_temp, bl); | |
888 | decode(new_primary_temp, bl); | |
7c673cae | 889 | if (struct_v >= 2) |
11fdf7f2 | 890 | decode(new_primary_affinity, bl); |
7c673cae FG |
891 | else |
892 | new_primary_affinity.clear(); | |
893 | if (struct_v >= 3) { | |
11fdf7f2 TL |
894 | decode(new_erasure_code_profiles, bl); |
895 | decode(old_erasure_code_profiles, bl); | |
7c673cae FG |
896 | } else { |
897 | new_erasure_code_profiles.clear(); | |
898 | old_erasure_code_profiles.clear(); | |
899 | } | |
900 | if (struct_v >= 4) { | |
11fdf7f2 TL |
901 | decode(new_pg_upmap, bl); |
902 | decode(old_pg_upmap, bl); | |
903 | decode(new_pg_upmap_items, bl); | |
904 | decode(old_pg_upmap_items, bl); | |
905 | } | |
906 | if (struct_v >= 6) { | |
907 | decode(new_removed_snaps, bl); | |
908 | decode(new_purged_snaps, bl); | |
909 | } | |
910 | if (struct_v >= 8) { | |
911 | decode(new_last_up_change, bl); | |
912 | decode(new_last_in_change, bl); | |
7c673cae FG |
913 | } |
914 | DECODE_FINISH(bl); // client-usable data | |
915 | } | |
916 | ||
917 | { | |
f67539c2 | 918 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
919 | decode(new_hb_back_up, bl); |
920 | decode(new_up_thru, bl); | |
921 | decode(new_last_clean_interval, bl); | |
922 | decode(new_lost, bl); | |
f67539c2 TL |
923 | decode(new_blocklist, bl); |
924 | decode(old_blocklist, bl); | |
11fdf7f2 TL |
925 | decode(new_up_cluster, bl); |
926 | decode(cluster_snapshot, bl); | |
927 | decode(new_uuid, bl); | |
928 | decode(new_xinfo, bl); | |
929 | decode(new_hb_front_up, bl); | |
7c673cae | 930 | if (struct_v >= 2) |
11fdf7f2 | 931 | decode(encode_features, bl); |
7c673cae FG |
932 | else |
933 | encode_features = CEPH_FEATURE_PGID64 | CEPH_FEATURE_OSDMAP_ENC; | |
934 | if (struct_v >= 3) { | |
11fdf7f2 TL |
935 | decode(new_nearfull_ratio, bl); |
936 | decode(new_full_ratio, bl); | |
7c673cae FG |
937 | } else { |
938 | new_nearfull_ratio = -1; | |
939 | new_full_ratio = -1; | |
940 | } | |
941 | if (struct_v >= 4) { | |
11fdf7f2 | 942 | decode(new_backfillfull_ratio, bl); |
7c673cae FG |
943 | } else { |
944 | new_backfillfull_ratio = -1; | |
945 | } | |
31f18b77 FG |
946 | if (struct_v == 5) { |
947 | string r; | |
11fdf7f2 | 948 | decode(r, bl); |
31f18b77 | 949 | if (r.length()) { |
9f95a23c | 950 | new_require_min_compat_client = ceph_release_from_name(r); |
31f18b77 FG |
951 | } |
952 | } | |
953 | if (struct_v >= 6) { | |
11fdf7f2 TL |
954 | decode(new_require_min_compat_client, bl); |
955 | decode(new_require_osd_release, bl); | |
31f18b77 FG |
956 | } else { |
957 | if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_LUMINOUS)) { | |
958 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 959 | new_require_osd_release = ceph_release_t::luminous; |
31f18b77 FG |
960 | new_flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
961 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_KRAKEN)) { | |
9f95a23c | 962 | new_require_osd_release = ceph_release_t::kraken; |
31f18b77 | 963 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_JEWEL)) { |
9f95a23c | 964 | new_require_osd_release = ceph_release_t::jewel; |
31f18b77 | 965 | } else { |
9f95a23c | 966 | new_require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
967 | } |
968 | } | |
81eedcae TL |
969 | if (struct_v >= 8) { |
970 | decode(new_crush_node_flags, bl); | |
971 | } | |
972 | if (struct_v >= 9) { | |
973 | decode(new_device_class_flags, bl); | |
974 | } | |
f67539c2 TL |
975 | if (struct_v >= 10) { |
976 | decode(change_stretch_mode, bl); | |
977 | decode(new_stretch_bucket_count, bl); | |
978 | decode(new_degraded_stretch_mode, bl); | |
979 | decode(new_recovering_stretch_mode, bl); | |
980 | decode(new_stretch_mode_bucket, bl); | |
981 | decode(stretch_mode_enabled, bl); | |
982 | } | |
983 | ||
7c673cae FG |
984 | DECODE_FINISH(bl); // osd-only data |
985 | } | |
986 | ||
987 | if (struct_v >= 8) { | |
988 | have_crc = true; | |
989 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 990 | decode(inc_crc, bl); |
7c673cae | 991 | tail_offset = bl.get_off(); |
11fdf7f2 | 992 | decode(full_crc, bl); |
7c673cae FG |
993 | } else { |
994 | have_crc = false; | |
995 | full_crc = 0; | |
996 | inc_crc = 0; | |
997 | } | |
998 | ||
999 | DECODE_FINISH(bl); // wrapper | |
1000 | ||
1001 | if (have_crc) { | |
1002 | // verify crc | |
1003 | uint32_t actual = crc_front.crc32c(-1); | |
1004 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 1005 | ceph::buffer::list tail; |
7c673cae FG |
1006 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
1007 | actual = tail.crc32c(actual); | |
1008 | } | |
1009 | if (inc_crc != actual) { | |
1010 | ostringstream ss; | |
1011 | ss << "bad crc, actual " << actual << " != expected " << inc_crc; | |
1012 | string s = ss.str(); | |
9f95a23c | 1013 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
1014 | } |
1015 | } | |
1016 | } | |
1017 | ||
1018 | void OSDMap::Incremental::dump(Formatter *f) const | |
1019 | { | |
1020 | f->dump_int("epoch", epoch); | |
1021 | f->dump_stream("fsid") << fsid; | |
1022 | f->dump_stream("modified") << modified; | |
11fdf7f2 TL |
1023 | f->dump_stream("new_last_up_change") << new_last_up_change; |
1024 | f->dump_stream("new_last_in_change") << new_last_in_change; | |
7c673cae FG |
1025 | f->dump_int("new_pool_max", new_pool_max); |
1026 | f->dump_int("new_flags", new_flags); | |
1027 | f->dump_float("new_full_ratio", new_full_ratio); | |
1028 | f->dump_float("new_nearfull_ratio", new_nearfull_ratio); | |
1029 | f->dump_float("new_backfillfull_ratio", new_backfillfull_ratio); | |
f67539c2 TL |
1030 | f->dump_int("new_require_min_compat_client", to_integer<int>(new_require_min_compat_client)); |
1031 | f->dump_int("new_require_osd_release", to_integer<int>(new_require_osd_release)); | |
7c673cae FG |
1032 | |
1033 | if (fullmap.length()) { | |
1034 | f->open_object_section("full_map"); | |
1035 | OSDMap full; | |
9f95a23c | 1036 | ceph::buffer::list fbl = fullmap; // kludge around constness. |
11fdf7f2 | 1037 | auto p = fbl.cbegin(); |
7c673cae FG |
1038 | full.decode(p); |
1039 | full.dump(f); | |
1040 | f->close_section(); | |
1041 | } | |
1042 | if (crush.length()) { | |
1043 | f->open_object_section("crush"); | |
1044 | CrushWrapper c; | |
9f95a23c | 1045 | ceph::buffer::list tbl = crush; // kludge around constness. |
11fdf7f2 | 1046 | auto p = tbl.cbegin(); |
7c673cae FG |
1047 | c.decode(p); |
1048 | c.dump(f); | |
1049 | f->close_section(); | |
1050 | } | |
1051 | ||
1052 | f->dump_int("new_max_osd", new_max_osd); | |
1053 | ||
1054 | f->open_array_section("new_pools"); | |
1055 | ||
1056 | for (const auto &new_pool : new_pools) { | |
1057 | f->open_object_section("pool"); | |
1058 | f->dump_int("pool", new_pool.first); | |
1059 | new_pool.second.dump(f); | |
1060 | f->close_section(); | |
1061 | } | |
1062 | f->close_section(); | |
1063 | f->open_array_section("new_pool_names"); | |
1064 | ||
1065 | for (const auto &new_pool_name : new_pool_names) { | |
1066 | f->open_object_section("pool_name"); | |
1067 | f->dump_int("pool", new_pool_name.first); | |
1068 | f->dump_string("name", new_pool_name.second); | |
1069 | f->close_section(); | |
1070 | } | |
1071 | f->close_section(); | |
1072 | f->open_array_section("old_pools"); | |
1073 | ||
1074 | for (const auto &old_pool : old_pools) | |
1075 | f->dump_int("pool", old_pool); | |
1076 | f->close_section(); | |
1077 | ||
1078 | f->open_array_section("new_up_osds"); | |
1079 | ||
1080 | for (const auto &upclient : new_up_client) { | |
1081 | f->open_object_section("osd"); | |
1082 | f->dump_int("osd", upclient.first); | |
11fdf7f2 TL |
1083 | f->dump_stream("public_addr") << upclient.second.legacy_addr(); |
1084 | f->dump_object("public_addrs", upclient.second); | |
1085 | if (auto p = new_up_cluster.find(upclient.first); | |
1086 | p != new_up_cluster.end()) { | |
1087 | f->dump_stream("cluster_addr") << p->second.legacy_addr(); | |
1088 | f->dump_object("cluster_addrs", p->second); | |
1089 | } | |
1090 | if (auto p = new_hb_back_up.find(upclient.first); | |
1091 | p != new_hb_back_up.end()) { | |
1092 | f->dump_object("heartbeat_back_addrs", p->second); | |
1093 | } | |
1094 | if (auto p = new_hb_front_up.find(upclient.first); | |
1095 | p != new_hb_front_up.end()) { | |
1096 | f->dump_object("heartbeat_front_addrs", p->second); | |
1097 | } | |
7c673cae FG |
1098 | f->close_section(); |
1099 | } | |
1100 | f->close_section(); | |
1101 | ||
1102 | f->open_array_section("new_weight"); | |
1103 | ||
1104 | for (const auto &weight : new_weight) { | |
1105 | f->open_object_section("osd"); | |
1106 | f->dump_int("osd", weight.first); | |
1107 | f->dump_int("weight", weight.second); | |
1108 | f->close_section(); | |
1109 | } | |
1110 | f->close_section(); | |
1111 | ||
1112 | f->open_array_section("osd_state_xor"); | |
1113 | for (const auto &ns : new_state) { | |
1114 | f->open_object_section("osd"); | |
1115 | f->dump_int("osd", ns.first); | |
1116 | set<string> st; | |
1117 | calc_state_set(new_state.find(ns.first)->second, st); | |
1118 | f->open_array_section("state_xor"); | |
1119 | for (auto &state : st) | |
1120 | f->dump_string("state", state); | |
1121 | f->close_section(); | |
c07f9fc5 | 1122 | f->close_section(); |
7c673cae FG |
1123 | } |
1124 | f->close_section(); | |
1125 | ||
1126 | f->open_array_section("new_pg_temp"); | |
1127 | ||
1128 | for (const auto &pg_temp : new_pg_temp) { | |
1129 | f->open_object_section("pg"); | |
1130 | f->dump_stream("pgid") << pg_temp.first; | |
1131 | f->open_array_section("osds"); | |
1132 | ||
1133 | for (const auto &osd : pg_temp.second) | |
1134 | f->dump_int("osd", osd); | |
1135 | f->close_section(); | |
1136 | f->close_section(); | |
1137 | } | |
1138 | f->close_section(); | |
1139 | ||
1140 | f->open_array_section("primary_temp"); | |
1141 | ||
1142 | for (const auto &primary_temp : new_primary_temp) { | |
1143 | f->dump_stream("pgid") << primary_temp.first; | |
1144 | f->dump_int("osd", primary_temp.second); | |
1145 | } | |
1146 | f->close_section(); // primary_temp | |
1147 | ||
1148 | f->open_array_section("new_pg_upmap"); | |
1149 | for (auto& i : new_pg_upmap) { | |
1150 | f->open_object_section("mapping"); | |
1151 | f->dump_stream("pgid") << i.first; | |
1152 | f->open_array_section("osds"); | |
1153 | for (auto osd : i.second) { | |
1154 | f->dump_int("osd", osd); | |
1155 | } | |
1156 | f->close_section(); | |
1157 | f->close_section(); | |
1158 | } | |
1159 | f->close_section(); | |
1160 | f->open_array_section("old_pg_upmap"); | |
1161 | for (auto& i : old_pg_upmap) { | |
1162 | f->dump_stream("pgid") << i; | |
1163 | } | |
1164 | f->close_section(); | |
1165 | ||
1166 | f->open_array_section("new_pg_upmap_items"); | |
1167 | for (auto& i : new_pg_upmap_items) { | |
1168 | f->open_object_section("mapping"); | |
1169 | f->dump_stream("pgid") << i.first; | |
1170 | f->open_array_section("mappings"); | |
1171 | for (auto& p : i.second) { | |
1172 | f->open_object_section("mapping"); | |
1173 | f->dump_int("from", p.first); | |
1174 | f->dump_int("to", p.second); | |
1175 | f->close_section(); | |
1176 | } | |
1177 | f->close_section(); | |
1178 | f->close_section(); | |
1179 | } | |
1180 | f->close_section(); | |
1181 | f->open_array_section("old_pg_upmap_items"); | |
1182 | for (auto& i : old_pg_upmap_items) { | |
1183 | f->dump_stream("pgid") << i; | |
1184 | } | |
1185 | f->close_section(); | |
1186 | ||
1187 | f->open_array_section("new_up_thru"); | |
1188 | ||
1189 | for (const auto &up_thru : new_up_thru) { | |
1190 | f->open_object_section("osd"); | |
1191 | f->dump_int("osd", up_thru.first); | |
1192 | f->dump_int("up_thru", up_thru.second); | |
1193 | f->close_section(); | |
1194 | } | |
1195 | f->close_section(); | |
1196 | ||
1197 | f->open_array_section("new_lost"); | |
1198 | ||
1199 | for (const auto &lost : new_lost) { | |
1200 | f->open_object_section("osd"); | |
1201 | f->dump_int("osd", lost.first); | |
1202 | f->dump_int("epoch_lost", lost.second); | |
1203 | f->close_section(); | |
1204 | } | |
1205 | f->close_section(); | |
1206 | ||
1207 | f->open_array_section("new_last_clean_interval"); | |
1208 | ||
1209 | for (const auto &last_clean_interval : new_last_clean_interval) { | |
1210 | f->open_object_section("osd"); | |
1211 | f->dump_int("osd", last_clean_interval.first); | |
1212 | f->dump_int("first", last_clean_interval.second.first); | |
1213 | f->dump_int("last", last_clean_interval.second.second); | |
1214 | f->close_section(); | |
1215 | } | |
1216 | f->close_section(); | |
1217 | ||
f67539c2 TL |
1218 | f->open_array_section("new_blocklist"); |
1219 | for (const auto &blist : new_blocklist) { | |
7c673cae FG |
1220 | stringstream ss; |
1221 | ss << blist.first; | |
1222 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1223 | } | |
1224 | f->close_section(); | |
f67539c2 TL |
1225 | f->open_array_section("old_blocklist"); |
1226 | for (const auto &blist : old_blocklist) | |
7c673cae FG |
1227 | f->dump_stream("addr") << blist; |
1228 | f->close_section(); | |
1229 | ||
1230 | f->open_array_section("new_xinfo"); | |
1231 | for (const auto &xinfo : new_xinfo) { | |
1232 | f->open_object_section("xinfo"); | |
1233 | f->dump_int("osd", xinfo.first); | |
1234 | xinfo.second.dump(f); | |
1235 | f->close_section(); | |
1236 | } | |
1237 | f->close_section(); | |
1238 | ||
1239 | if (cluster_snapshot.size()) | |
1240 | f->dump_string("cluster_snapshot", cluster_snapshot); | |
1241 | ||
1242 | f->open_array_section("new_uuid"); | |
1243 | for (const auto &uuid : new_uuid) { | |
1244 | f->open_object_section("osd"); | |
1245 | f->dump_int("osd", uuid.first); | |
1246 | f->dump_stream("uuid") << uuid.second; | |
1247 | f->close_section(); | |
1248 | } | |
1249 | f->close_section(); | |
1250 | ||
1251 | OSDMap::dump_erasure_code_profiles(new_erasure_code_profiles, f); | |
1252 | f->open_array_section("old_erasure_code_profiles"); | |
1253 | for (const auto &erasure_code_profile : old_erasure_code_profiles) { | |
9f95a23c | 1254 | f->dump_string("old", erasure_code_profile); |
7c673cae FG |
1255 | } |
1256 | f->close_section(); | |
11fdf7f2 TL |
1257 | |
1258 | f->open_array_section("new_removed_snaps"); | |
1259 | for (auto& p : new_removed_snaps) { | |
1260 | f->open_object_section("pool"); | |
1261 | f->dump_int("pool", p.first); | |
1262 | f->open_array_section("snaps"); | |
1263 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1264 | f->open_object_section("interval"); | |
1265 | f->dump_unsigned("begin", q.get_start()); | |
1266 | f->dump_unsigned("length", q.get_len()); | |
1267 | f->close_section(); | |
1268 | } | |
1269 | f->close_section(); | |
1270 | f->close_section(); | |
1271 | } | |
1272 | f->close_section(); | |
1273 | f->open_array_section("new_purged_snaps"); | |
1274 | for (auto& p : new_purged_snaps) { | |
1275 | f->open_object_section("pool"); | |
1276 | f->dump_int("pool", p.first); | |
1277 | f->open_array_section("snaps"); | |
1278 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1279 | f->open_object_section("interval"); | |
1280 | f->dump_unsigned("begin", q.get_start()); | |
1281 | f->dump_unsigned("length", q.get_len()); | |
1282 | f->close_section(); | |
1283 | } | |
1284 | f->close_section(); | |
1285 | f->close_section(); | |
1286 | } | |
81eedcae TL |
1287 | f->open_array_section("new_crush_node_flags"); |
1288 | for (auto& i : new_crush_node_flags) { | |
1289 | f->open_object_section("node"); | |
1290 | f->dump_int("id", i.first); | |
1291 | set<string> st; | |
1292 | calc_state_set(i.second, st); | |
1293 | for (auto& j : st) { | |
1294 | f->dump_string("flag", j); | |
1295 | } | |
1296 | f->close_section(); | |
1297 | } | |
1298 | f->close_section(); | |
1299 | f->open_array_section("new_device_class_flags"); | |
1300 | for (auto& i : new_device_class_flags) { | |
1301 | f->open_object_section("device_class"); | |
1302 | f->dump_int("id", i.first); | |
1303 | set<string> st; | |
1304 | calc_state_set(i.second, st); | |
1305 | for (auto& j : st) { | |
1306 | f->dump_string("flag", j); | |
1307 | } | |
1308 | f->close_section(); | |
1309 | } | |
1310 | f->close_section(); | |
f67539c2 TL |
1311 | f->open_object_section("stretch_mode"); |
1312 | { | |
1313 | f->dump_bool("change_stretch_mode", change_stretch_mode); | |
1314 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
1315 | f->dump_unsigned("new_stretch_bucket_count", new_stretch_bucket_count); | |
1316 | f->dump_unsigned("new_degraded_stretch_mode", new_degraded_stretch_mode); | |
1317 | f->dump_unsigned("new_recovering_stretch_mode", new_recovering_stretch_mode); | |
1318 | f->dump_int("new_stretch_mode_bucket", new_stretch_mode_bucket); | |
1319 | } | |
1320 | f->close_section(); | |
11fdf7f2 | 1321 | f->close_section(); |
7c673cae FG |
1322 | } |
1323 | ||
1324 | void OSDMap::Incremental::generate_test_instances(list<Incremental*>& o) | |
1325 | { | |
1326 | o.push_back(new Incremental); | |
1327 | } | |
1328 | ||
1329 | // ---------------------------------- | |
1330 | // OSDMap | |
1331 | ||
1332 | void OSDMap::set_epoch(epoch_t e) | |
1333 | { | |
1334 | epoch = e; | |
1335 | for (auto &pool : pools) | |
1336 | pool.second.last_change = e; | |
1337 | } | |
1338 | ||
f67539c2 | 1339 | bool OSDMap::is_blocklisted(const entity_addr_t& orig) const |
7c673cae | 1340 | { |
f67539c2 | 1341 | if (blocklist.empty()) { |
7c673cae | 1342 | return false; |
11fdf7f2 TL |
1343 | } |
1344 | ||
f67539c2 | 1345 | // all blocklist entries are type ANY for nautilus+ |
11fdf7f2 TL |
1346 | // FIXME: avoid this copy! |
1347 | entity_addr_t a = orig; | |
9f95a23c | 1348 | if (require_osd_release < ceph_release_t::nautilus) { |
11fdf7f2 TL |
1349 | a.set_type(entity_addr_t::TYPE_LEGACY); |
1350 | } else { | |
1351 | a.set_type(entity_addr_t::TYPE_ANY); | |
1352 | } | |
7c673cae FG |
1353 | |
1354 | // this specific instance? | |
f67539c2 | 1355 | if (blocklist.count(a)) { |
7c673cae | 1356 | return true; |
11fdf7f2 | 1357 | } |
7c673cae | 1358 | |
f67539c2 | 1359 | // is entire ip blocklisted? |
7c673cae | 1360 | if (a.is_ip()) { |
11fdf7f2 TL |
1361 | a.set_port(0); |
1362 | a.set_nonce(0); | |
f67539c2 | 1363 | if (blocklist.count(a)) { |
11fdf7f2 TL |
1364 | return true; |
1365 | } | |
1366 | } | |
1367 | ||
1368 | return false; | |
1369 | } | |
1370 | ||
f67539c2 | 1371 | bool OSDMap::is_blocklisted(const entity_addrvec_t& av) const |
11fdf7f2 | 1372 | { |
f67539c2 | 1373 | if (blocklist.empty()) |
11fdf7f2 TL |
1374 | return false; |
1375 | ||
1376 | for (auto& a : av.v) { | |
f67539c2 | 1377 | if (is_blocklisted(a)) { |
7c673cae FG |
1378 | return true; |
1379 | } | |
1380 | } | |
1381 | ||
1382 | return false; | |
1383 | } | |
1384 | ||
f67539c2 | 1385 | void OSDMap::get_blocklist(list<pair<entity_addr_t,utime_t> > *bl) const |
7c673cae | 1386 | { |
f67539c2 | 1387 | std::copy(blocklist.begin(), blocklist.end(), std::back_inserter(*bl)); |
7c673cae FG |
1388 | } |
1389 | ||
f67539c2 | 1390 | void OSDMap::get_blocklist(std::set<entity_addr_t> *bl) const |
31f18b77 | 1391 | { |
f67539c2 | 1392 | for (const auto &i : blocklist) { |
31f18b77 FG |
1393 | bl->insert(i.first); |
1394 | } | |
1395 | } | |
1396 | ||
7c673cae FG |
1397 | void OSDMap::set_max_osd(int m) |
1398 | { | |
7c673cae | 1399 | max_osd = m; |
f67539c2 TL |
1400 | osd_state.resize(max_osd, 0); |
1401 | osd_weight.resize(max_osd, CEPH_OSD_OUT); | |
1402 | osd_info.resize(max_osd); | |
1403 | osd_xinfo.resize(max_osd); | |
1404 | osd_addrs->client_addrs.resize(max_osd); | |
1405 | osd_addrs->cluster_addrs.resize(max_osd); | |
1406 | osd_addrs->hb_back_addrs.resize(max_osd); | |
1407 | osd_addrs->hb_front_addrs.resize(max_osd); | |
1408 | osd_uuid->resize(max_osd); | |
7c673cae | 1409 | if (osd_primary_affinity) |
f67539c2 | 1410 | osd_primary_affinity->resize(max_osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); |
7c673cae FG |
1411 | |
1412 | calc_num_osds(); | |
1413 | } | |
1414 | ||
1415 | int OSDMap::calc_num_osds() | |
1416 | { | |
1417 | num_osd = 0; | |
1418 | num_up_osd = 0; | |
1419 | num_in_osd = 0; | |
1420 | for (int i=0; i<max_osd; i++) { | |
1421 | if (osd_state[i] & CEPH_OSD_EXISTS) { | |
1422 | ++num_osd; | |
1423 | if (osd_state[i] & CEPH_OSD_UP) { | |
1424 | ++num_up_osd; | |
1425 | } | |
1426 | if (get_weight(i) != CEPH_OSD_OUT) { | |
1427 | ++num_in_osd; | |
1428 | } | |
1429 | } | |
1430 | } | |
1431 | return num_osd; | |
1432 | } | |
1433 | ||
3efd9988 FG |
1434 | void OSDMap::get_full_pools(CephContext *cct, |
1435 | set<int64_t> *full, | |
1436 | set<int64_t> *backfillfull, | |
1437 | set<int64_t> *nearfull) const | |
7c673cae | 1438 | { |
11fdf7f2 TL |
1439 | ceph_assert(full); |
1440 | ceph_assert(backfillfull); | |
1441 | ceph_assert(nearfull); | |
3efd9988 FG |
1442 | full->clear(); |
1443 | backfillfull->clear(); | |
1444 | nearfull->clear(); | |
1445 | ||
1446 | vector<int> full_osds; | |
1447 | vector<int> backfillfull_osds; | |
1448 | vector<int> nearfull_osds; | |
7c673cae FG |
1449 | for (int i = 0; i < max_osd; ++i) { |
1450 | if (exists(i) && is_up(i) && is_in(i)) { | |
1451 | if (osd_state[i] & CEPH_OSD_FULL) | |
3efd9988 | 1452 | full_osds.push_back(i); |
7c673cae | 1453 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) |
3efd9988 | 1454 | backfillfull_osds.push_back(i); |
7c673cae | 1455 | else if (osd_state[i] & CEPH_OSD_NEARFULL) |
3efd9988 | 1456 | nearfull_osds.push_back(i); |
7c673cae FG |
1457 | } |
1458 | } | |
3efd9988 FG |
1459 | |
1460 | for (auto i: full_osds) { | |
1461 | get_pool_ids_by_osd(cct, i, full); | |
1462 | } | |
1463 | for (auto i: backfillfull_osds) { | |
1464 | get_pool_ids_by_osd(cct, i, backfillfull); | |
1465 | } | |
1466 | for (auto i: nearfull_osds) { | |
1467 | get_pool_ids_by_osd(cct, i, nearfull); | |
1468 | } | |
7c673cae FG |
1469 | } |
1470 | ||
31f18b77 FG |
1471 | void OSDMap::get_full_osd_counts(set<int> *full, set<int> *backfill, |
1472 | set<int> *nearfull) const | |
1473 | { | |
1474 | full->clear(); | |
1475 | backfill->clear(); | |
1476 | nearfull->clear(); | |
1477 | for (int i = 0; i < max_osd; ++i) { | |
1478 | if (exists(i) && is_up(i) && is_in(i)) { | |
1479 | if (osd_state[i] & CEPH_OSD_FULL) | |
1480 | full->emplace(i); | |
1481 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1482 | backfill->emplace(i); | |
1483 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1484 | nearfull->emplace(i); | |
1485 | } | |
1486 | } | |
1487 | } | |
1488 | ||
7c673cae FG |
1489 | void OSDMap::get_all_osds(set<int32_t>& ls) const |
1490 | { | |
1491 | for (int i=0; i<max_osd; i++) | |
1492 | if (exists(i)) | |
1493 | ls.insert(i); | |
1494 | } | |
1495 | ||
1496 | void OSDMap::get_up_osds(set<int32_t>& ls) const | |
1497 | { | |
1498 | for (int i = 0; i < max_osd; i++) { | |
1499 | if (is_up(i)) | |
1500 | ls.insert(i); | |
1501 | } | |
1502 | } | |
1503 | ||
81eedcae | 1504 | void OSDMap::get_out_existing_osds(set<int32_t>& ls) const |
31f18b77 FG |
1505 | { |
1506 | for (int i = 0; i < max_osd; i++) { | |
81eedcae | 1507 | if (exists(i) && get_weight(i) == CEPH_OSD_OUT) |
31f18b77 FG |
1508 | ls.insert(i); |
1509 | } | |
1510 | } | |
1511 | ||
11fdf7f2 TL |
1512 | void OSDMap::get_flag_set(set<string> *flagset) const |
1513 | { | |
1514 | for (unsigned i = 0; i < sizeof(flags) * 8; ++i) { | |
1515 | if (flags & (1<<i)) { | |
1516 | flagset->insert(get_flag_string(flags & (1<<i))); | |
1517 | } | |
1518 | } | |
1519 | } | |
1520 | ||
7c673cae FG |
1521 | void OSDMap::calc_state_set(int state, set<string>& st) |
1522 | { | |
1523 | unsigned t = state; | |
1524 | for (unsigned s = 1; t; s <<= 1) { | |
1525 | if (t & s) { | |
1526 | t &= ~s; | |
1527 | st.insert(ceph_osd_state_name(s)); | |
1528 | } | |
1529 | } | |
1530 | } | |
1531 | ||
1532 | void OSDMap::adjust_osd_weights(const map<int,double>& weights, Incremental& inc) const | |
1533 | { | |
1534 | float max = 0; | |
1535 | for (const auto &weight : weights) { | |
1536 | if (weight.second > max) | |
1537 | max = weight.second; | |
1538 | } | |
1539 | ||
1540 | for (const auto &weight : weights) { | |
1541 | inc.new_weight[weight.first] = (unsigned)((weight.second / max) * CEPH_OSD_IN); | |
1542 | } | |
1543 | } | |
1544 | ||
1545 | int OSDMap::identify_osd(const entity_addr_t& addr) const | |
1546 | { | |
1547 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1548 | if (exists(i) && (get_addrs(i).contains(addr) || |
1549 | get_cluster_addrs(i).contains(addr))) | |
7c673cae FG |
1550 | return i; |
1551 | return -1; | |
1552 | } | |
1553 | ||
1554 | int OSDMap::identify_osd(const uuid_d& u) const | |
1555 | { | |
1556 | for (int i=0; i<max_osd; i++) | |
1557 | if (exists(i) && get_uuid(i) == u) | |
1558 | return i; | |
1559 | return -1; | |
1560 | } | |
1561 | ||
1562 | int OSDMap::identify_osd_on_all_channels(const entity_addr_t& addr) const | |
1563 | { | |
1564 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1565 | if (exists(i) && (get_addrs(i).contains(addr) || |
1566 | get_cluster_addrs(i).contains(addr) || | |
1567 | get_hb_back_addrs(i).contains(addr) || | |
1568 | get_hb_front_addrs(i).contains(addr))) | |
7c673cae FG |
1569 | return i; |
1570 | return -1; | |
1571 | } | |
1572 | ||
1573 | int OSDMap::find_osd_on_ip(const entity_addr_t& ip) const | |
1574 | { | |
1575 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1576 | if (exists(i) && (get_addrs(i).is_same_host(ip) || |
1577 | get_cluster_addrs(i).is_same_host(ip))) | |
7c673cae FG |
1578 | return i; |
1579 | return -1; | |
1580 | } | |
1581 | ||
1582 | ||
1583 | uint64_t OSDMap::get_features(int entity_type, uint64_t *pmask) const | |
1584 | { | |
1585 | uint64_t features = 0; // things we actually have | |
1586 | uint64_t mask = 0; // things we could have | |
1587 | ||
1588 | if (crush->has_nondefault_tunables()) | |
1589 | features |= CEPH_FEATURE_CRUSH_TUNABLES; | |
1590 | if (crush->has_nondefault_tunables2()) | |
1591 | features |= CEPH_FEATURE_CRUSH_TUNABLES2; | |
1592 | if (crush->has_nondefault_tunables3()) | |
1593 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1594 | if (crush->has_v4_buckets()) | |
1595 | features |= CEPH_FEATURE_CRUSH_V4; | |
1596 | if (crush->has_nondefault_tunables5()) | |
1597 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
c07f9fc5 FG |
1598 | if (crush->has_incompat_choose_args()) { |
1599 | features |= CEPH_FEATUREMASK_CRUSH_CHOOSE_ARGS; | |
1600 | } | |
7c673cae FG |
1601 | mask |= CEPH_FEATURES_CRUSH; |
1602 | ||
1603 | if (!pg_upmap.empty() || !pg_upmap_items.empty()) | |
1604 | features |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1605 | mask |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1606 | ||
1607 | for (auto &pool: pools) { | |
1608 | if (pool.second.has_flag(pg_pool_t::FLAG_HASHPSPOOL)) { | |
1609 | features |= CEPH_FEATURE_OSDHASHPSPOOL; | |
1610 | } | |
7c673cae FG |
1611 | if (!pool.second.tiers.empty() || |
1612 | pool.second.is_tier()) { | |
1613 | features |= CEPH_FEATURE_OSD_CACHEPOOL; | |
1614 | } | |
20effc67 | 1615 | int ruleid = pool.second.get_crush_rule(); |
7c673cae FG |
1616 | if (ruleid >= 0) { |
1617 | if (crush->is_v2_rule(ruleid)) | |
1618 | features |= CEPH_FEATURE_CRUSH_V2; | |
1619 | if (crush->is_v3_rule(ruleid)) | |
1620 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1621 | if (crush->is_v5_rule(ruleid)) | |
1622 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
1623 | } | |
1624 | } | |
7c673cae | 1625 | mask |= CEPH_FEATURE_OSDHASHPSPOOL | CEPH_FEATURE_OSD_CACHEPOOL; |
7c673cae FG |
1626 | |
1627 | if (osd_primary_affinity) { | |
1628 | for (int i = 0; i < max_osd; ++i) { | |
1629 | if ((*osd_primary_affinity)[i] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
1630 | features |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1631 | break; | |
1632 | } | |
1633 | } | |
1634 | } | |
1635 | mask |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1636 | ||
1637 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1638 | const uint64_t jewel_features = CEPH_FEATURE_SERVER_JEWEL; | |
9f95a23c | 1639 | if (require_osd_release >= ceph_release_t::jewel) { |
7c673cae FG |
1640 | features |= jewel_features; |
1641 | } | |
1642 | mask |= jewel_features; | |
1643 | ||
1644 | const uint64_t kraken_features = CEPH_FEATUREMASK_SERVER_KRAKEN | |
1645 | | CEPH_FEATURE_MSG_ADDR2; | |
9f95a23c | 1646 | if (require_osd_release >= ceph_release_t::kraken) { |
7c673cae FG |
1647 | features |= kraken_features; |
1648 | } | |
1649 | mask |= kraken_features; | |
f67539c2 TL |
1650 | |
1651 | if (stretch_mode_enabled) { | |
1652 | features |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1653 | mask |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1654 | } | |
7c673cae FG |
1655 | } |
1656 | ||
9f95a23c | 1657 | if (require_min_compat_client >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
1658 | // if min_compat_client is >= nautilus, require v2 cephx signatures |
1659 | // from everyone | |
1660 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
9f95a23c | 1661 | } else if (require_osd_release >= ceph_release_t::nautilus && |
11fdf7f2 TL |
1662 | entity_type == CEPH_ENTITY_TYPE_OSD) { |
1663 | // if osds are >= nautilus, at least require the signatures from them | |
1664 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
1665 | } | |
1666 | mask |= CEPH_FEATUREMASK_CEPHX_V2; | |
1667 | ||
7c673cae FG |
1668 | if (pmask) |
1669 | *pmask = mask; | |
1670 | return features; | |
1671 | } | |
1672 | ||
9f95a23c | 1673 | ceph_release_t OSDMap::get_min_compat_client() const |
7c673cae FG |
1674 | { |
1675 | uint64_t f = get_features(CEPH_ENTITY_TYPE_CLIENT, nullptr); | |
1676 | ||
1677 | if (HAVE_FEATURE(f, OSDMAP_PG_UPMAP) || // v12.0.0-1733-g27d6f43 | |
31f18b77 | 1678 | HAVE_FEATURE(f, CRUSH_CHOOSE_ARGS)) { // v12.0.1-2172-gef1ef28 |
9f95a23c | 1679 | return ceph_release_t::luminous; // v12.2.0 |
7c673cae FG |
1680 | } |
1681 | if (HAVE_FEATURE(f, CRUSH_TUNABLES5)) { // v10.0.0-612-g043a737 | |
9f95a23c | 1682 | return ceph_release_t::jewel; // v10.2.0 |
7c673cae FG |
1683 | } |
1684 | if (HAVE_FEATURE(f, CRUSH_V4)) { // v0.91-678-g325fc56 | |
9f95a23c | 1685 | return ceph_release_t::hammer; // v0.94.0 |
7c673cae FG |
1686 | } |
1687 | if (HAVE_FEATURE(f, OSD_PRIMARY_AFFINITY) || // v0.76-553-gf825624 | |
1688 | HAVE_FEATURE(f, CRUSH_TUNABLES3) || // v0.76-395-ge20a55d | |
7c673cae | 1689 | HAVE_FEATURE(f, OSD_CACHEPOOL)) { // v0.67-401-gb91c1c5 |
9f95a23c | 1690 | return ceph_release_t::firefly; // v0.80.0 |
7c673cae FG |
1691 | } |
1692 | if (HAVE_FEATURE(f, CRUSH_TUNABLES2) || // v0.54-684-g0cc47ff | |
1693 | HAVE_FEATURE(f, OSDHASHPSPOOL)) { // v0.57-398-g8cc2b0f | |
9f95a23c | 1694 | return ceph_release_t::dumpling; // v0.67.0 |
7c673cae FG |
1695 | } |
1696 | if (HAVE_FEATURE(f, CRUSH_TUNABLES)) { // v0.48argonaut-206-g6f381af | |
9f95a23c | 1697 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae | 1698 | } |
9f95a23c | 1699 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae FG |
1700 | } |
1701 | ||
9f95a23c | 1702 | ceph_release_t OSDMap::get_require_min_compat_client() const |
11fdf7f2 TL |
1703 | { |
1704 | return require_min_compat_client; | |
1705 | } | |
1706 | ||
7c673cae FG |
1707 | void OSDMap::_calc_up_osd_features() |
1708 | { | |
1709 | bool first = true; | |
1710 | cached_up_osd_features = 0; | |
1711 | for (int osd = 0; osd < max_osd; ++osd) { | |
1712 | if (!is_up(osd)) | |
1713 | continue; | |
1714 | const osd_xinfo_t &xi = get_xinfo(osd); | |
3efd9988 FG |
1715 | if (xi.features == 0) |
1716 | continue; // bogus xinfo, maybe #20751 or similar, skipping | |
7c673cae FG |
1717 | if (first) { |
1718 | cached_up_osd_features = xi.features; | |
1719 | first = false; | |
1720 | } else { | |
1721 | cached_up_osd_features &= xi.features; | |
1722 | } | |
1723 | } | |
1724 | } | |
1725 | ||
1726 | uint64_t OSDMap::get_up_osd_features() const | |
1727 | { | |
1728 | return cached_up_osd_features; | |
1729 | } | |
1730 | ||
1731 | void OSDMap::dedup(const OSDMap *o, OSDMap *n) | |
1732 | { | |
11fdf7f2 | 1733 | using ceph::encode; |
7c673cae FG |
1734 | if (o->epoch == n->epoch) |
1735 | return; | |
1736 | ||
1737 | int diff = 0; | |
1738 | ||
1739 | // do addrs match? | |
1740 | if (o->max_osd != n->max_osd) | |
1741 | diff++; | |
1742 | for (int i = 0; i < o->max_osd && i < n->max_osd; i++) { | |
11fdf7f2 TL |
1743 | if ( n->osd_addrs->client_addrs[i] && o->osd_addrs->client_addrs[i] && |
1744 | *n->osd_addrs->client_addrs[i] == *o->osd_addrs->client_addrs[i]) | |
1745 | n->osd_addrs->client_addrs[i] = o->osd_addrs->client_addrs[i]; | |
7c673cae FG |
1746 | else |
1747 | diff++; | |
11fdf7f2 TL |
1748 | if ( n->osd_addrs->cluster_addrs[i] && o->osd_addrs->cluster_addrs[i] && |
1749 | *n->osd_addrs->cluster_addrs[i] == *o->osd_addrs->cluster_addrs[i]) | |
1750 | n->osd_addrs->cluster_addrs[i] = o->osd_addrs->cluster_addrs[i]; | |
7c673cae FG |
1751 | else |
1752 | diff++; | |
11fdf7f2 TL |
1753 | if ( n->osd_addrs->hb_back_addrs[i] && o->osd_addrs->hb_back_addrs[i] && |
1754 | *n->osd_addrs->hb_back_addrs[i] == *o->osd_addrs->hb_back_addrs[i]) | |
1755 | n->osd_addrs->hb_back_addrs[i] = o->osd_addrs->hb_back_addrs[i]; | |
7c673cae FG |
1756 | else |
1757 | diff++; | |
11fdf7f2 TL |
1758 | if ( n->osd_addrs->hb_front_addrs[i] && o->osd_addrs->hb_front_addrs[i] && |
1759 | *n->osd_addrs->hb_front_addrs[i] == *o->osd_addrs->hb_front_addrs[i]) | |
1760 | n->osd_addrs->hb_front_addrs[i] = o->osd_addrs->hb_front_addrs[i]; | |
7c673cae FG |
1761 | else |
1762 | diff++; | |
1763 | } | |
1764 | if (diff == 0) { | |
1765 | // zoinks, no differences at all! | |
1766 | n->osd_addrs = o->osd_addrs; | |
1767 | } | |
1768 | ||
1769 | // does crush match? | |
9f95a23c | 1770 | ceph::buffer::list oc, nc; |
11fdf7f2 TL |
1771 | encode(*o->crush, oc, CEPH_FEATURES_SUPPORTED_DEFAULT); |
1772 | encode(*n->crush, nc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
7c673cae FG |
1773 | if (oc.contents_equal(nc)) { |
1774 | n->crush = o->crush; | |
1775 | } | |
1776 | ||
1777 | // does pg_temp match? | |
31f18b77 FG |
1778 | if (*o->pg_temp == *n->pg_temp) |
1779 | n->pg_temp = o->pg_temp; | |
7c673cae FG |
1780 | |
1781 | // does primary_temp match? | |
1782 | if (o->primary_temp->size() == n->primary_temp->size()) { | |
1783 | if (*o->primary_temp == *n->primary_temp) | |
1784 | n->primary_temp = o->primary_temp; | |
1785 | } | |
1786 | ||
1787 | // do uuids match? | |
1788 | if (o->osd_uuid->size() == n->osd_uuid->size() && | |
1789 | *o->osd_uuid == *n->osd_uuid) | |
1790 | n->osd_uuid = o->osd_uuid; | |
1791 | } | |
1792 | ||
1793 | void OSDMap::clean_temps(CephContext *cct, | |
11fdf7f2 TL |
1794 | const OSDMap& oldmap, |
1795 | const OSDMap& nextmap, | |
1796 | Incremental *pending_inc) | |
7c673cae FG |
1797 | { |
1798 | ldout(cct, 10) << __func__ << dendl; | |
7c673cae | 1799 | |
11fdf7f2 | 1800 | for (auto pg : *nextmap.pg_temp) { |
7c673cae FG |
1801 | // if pool does not exist, remove any existing pg_temps associated with |
1802 | // it. we don't care about pg_temps on the pending_inc either; if there | |
1803 | // are new_pg_temp entries on the pending, clear them out just as well. | |
11fdf7f2 | 1804 | if (!nextmap.have_pg_pool(pg.first.pool())) { |
7c673cae FG |
1805 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first |
1806 | << " for nonexistent pool " << pg.first.pool() << dendl; | |
1807 | pending_inc->new_pg_temp[pg.first].clear(); | |
1808 | continue; | |
1809 | } | |
20effc67 TL |
1810 | if (!nextmap.pg_exists(pg.first)) { |
1811 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1812 | << " for nonexistent pg " << dendl; | |
1813 | pending_inc->new_pg_temp[pg.first].clear(); | |
1814 | continue; | |
1815 | } | |
7c673cae FG |
1816 | // all osds down? |
1817 | unsigned num_up = 0; | |
1818 | for (auto o : pg.second) { | |
11fdf7f2 | 1819 | if (!nextmap.is_down(o)) { |
7c673cae FG |
1820 | ++num_up; |
1821 | break; | |
1822 | } | |
1823 | } | |
1824 | if (num_up == 0) { | |
1825 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1826 | << " with all down osds" << pg.second << dendl; | |
1827 | pending_inc->new_pg_temp[pg.first].clear(); | |
1828 | continue; | |
1829 | } | |
1830 | // redundant pg_temp? | |
1831 | vector<int> raw_up; | |
1832 | int primary; | |
11fdf7f2 | 1833 | nextmap.pg_to_raw_up(pg.first, &raw_up, &primary); |
91327a77 | 1834 | bool remove = false; |
11fdf7f2 | 1835 | if (raw_up == pg.second) { |
7c673cae FG |
1836 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1837 | << pg.second << " that matches raw_up mapping" << dendl; | |
91327a77 AA |
1838 | remove = true; |
1839 | } | |
1840 | // oversized pg_temp? | |
11fdf7f2 | 1841 | if (pg.second.size() > nextmap.get_pg_pool(pg.first.pool())->get_size()) { |
91327a77 AA |
1842 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1843 | << pg.second << " exceeds pool size" << dendl; | |
1844 | remove = true; | |
1845 | } | |
1846 | if (remove) { | |
11fdf7f2 | 1847 | if (oldmap.pg_temp->count(pg.first)) |
7c673cae FG |
1848 | pending_inc->new_pg_temp[pg.first].clear(); |
1849 | else | |
1850 | pending_inc->new_pg_temp.erase(pg.first); | |
1851 | } | |
1852 | } | |
1853 | ||
11fdf7f2 | 1854 | for (auto &pg : *nextmap.primary_temp) { |
7c673cae | 1855 | // primary down? |
11fdf7f2 | 1856 | if (nextmap.is_down(pg.second)) { |
7c673cae FG |
1857 | ldout(cct, 10) << __func__ << " removing primary_temp " << pg.first |
1858 | << " to down " << pg.second << dendl; | |
1859 | pending_inc->new_primary_temp[pg.first] = -1; | |
1860 | continue; | |
1861 | } | |
1862 | // redundant primary_temp? | |
1863 | vector<int> real_up, templess_up; | |
1864 | int real_primary, templess_primary; | |
1865 | pg_t pgid = pg.first; | |
11fdf7f2 TL |
1866 | nextmap.pg_to_acting_osds(pgid, &real_up, &real_primary); |
1867 | nextmap.pg_to_raw_up(pgid, &templess_up, &templess_primary); | |
7c673cae FG |
1868 | if (real_primary == templess_primary){ |
1869 | ldout(cct, 10) << __func__ << " removing primary_temp " | |
1870 | << pgid << " -> " << real_primary | |
1871 | << " (unnecessary/redundant)" << dendl; | |
11fdf7f2 | 1872 | if (oldmap.primary_temp->count(pgid)) |
7c673cae FG |
1873 | pending_inc->new_primary_temp[pgid] = -1; |
1874 | else | |
1875 | pending_inc->new_primary_temp.erase(pgid); | |
1876 | } | |
1877 | } | |
1878 | } | |
1879 | ||
494da23a | 1880 | void OSDMap::get_upmap_pgs(vector<pg_t> *upmap_pgs) const |
94b18763 | 1881 | { |
494da23a TL |
1882 | upmap_pgs->reserve(pg_upmap.size() + pg_upmap_items.size()); |
1883 | for (auto& p : pg_upmap) | |
1884 | upmap_pgs->push_back(p.first); | |
1885 | for (auto& p : pg_upmap_items) | |
1886 | upmap_pgs->push_back(p.first); | |
1887 | } | |
94b18763 | 1888 | |
494da23a TL |
1889 | bool OSDMap::check_pg_upmaps( |
1890 | CephContext *cct, | |
1891 | const vector<pg_t>& to_check, | |
1892 | vector<pg_t> *to_cancel, | |
1893 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> *to_remap) const | |
1894 | { | |
1895 | bool any_change = false; | |
1896 | map<int, map<int, float>> rule_weight_map; | |
28e407b8 | 1897 | for (auto& pg : to_check) { |
494da23a | 1898 | const pg_pool_t *pi = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
1899 | if (!pi || pg.ps() >= pi->get_pg_num_pending()) { |
1900 | ldout(cct, 0) << __func__ << " pg " << pg << " is gone or merge source" | |
1901 | << dendl; | |
494da23a | 1902 | to_cancel->push_back(pg); |
11fdf7f2 TL |
1903 | continue; |
1904 | } | |
1905 | if (pi->is_pending_merge(pg, nullptr)) { | |
1906 | ldout(cct, 0) << __func__ << " pg " << pg << " is pending merge" | |
1907 | << dendl; | |
494da23a | 1908 | to_cancel->push_back(pg); |
94b18763 FG |
1909 | continue; |
1910 | } | |
494da23a TL |
1911 | vector<int> raw, up; |
1912 | pg_to_raw_upmap(pg, &raw, &up); | |
494da23a TL |
1913 | auto crush_rule = get_pg_pool_crush_rule(pg); |
1914 | auto r = crush->verify_upmap(cct, | |
1915 | crush_rule, | |
1916 | get_pg_pool_size(pg), | |
1917 | up); | |
a8e16298 TL |
1918 | if (r < 0) { |
1919 | ldout(cct, 0) << __func__ << " verify_upmap of pg " << pg | |
1920 | << " returning " << r | |
1921 | << dendl; | |
494da23a | 1922 | to_cancel->push_back(pg); |
a8e16298 TL |
1923 | continue; |
1924 | } | |
1925 | // below we check against crush-topology changing.. | |
28e407b8 AA |
1926 | map<int, float> weight_map; |
1927 | auto it = rule_weight_map.find(crush_rule); | |
1928 | if (it == rule_weight_map.end()) { | |
494da23a | 1929 | auto r = crush->get_rule_weight_osd_map(crush_rule, &weight_map); |
28e407b8 AA |
1930 | if (r < 0) { |
1931 | lderr(cct) << __func__ << " unable to get crush weight_map for " | |
494da23a TL |
1932 | << "crush_rule " << crush_rule |
1933 | << dendl; | |
28e407b8 AA |
1934 | continue; |
1935 | } | |
1936 | rule_weight_map[crush_rule] = weight_map; | |
1937 | } else { | |
1938 | weight_map = it->second; | |
1939 | } | |
28e407b8 | 1940 | ldout(cct, 10) << __func__ << " pg " << pg |
28e407b8 | 1941 | << " weight_map " << weight_map |
94b18763 | 1942 | << dendl; |
a8e16298 | 1943 | for (auto osd : up) { |
28e407b8 AA |
1944 | auto it = weight_map.find(osd); |
1945 | if (it == weight_map.end()) { | |
92f5a8d4 TL |
1946 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd << " is gone or has " |
1947 | << "been moved out of the specific crush-tree" | |
1948 | << dendl; | |
494da23a | 1949 | to_cancel->push_back(pg); |
94b18763 FG |
1950 | break; |
1951 | } | |
494da23a | 1952 | auto adjusted_weight = get_weightf(it->first) * it->second; |
28e407b8 | 1953 | if (adjusted_weight == 0) { |
92f5a8d4 TL |
1954 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd |
1955 | << " is out/crush-out" | |
1956 | << dendl; | |
494da23a | 1957 | to_cancel->push_back(pg); |
94b18763 FG |
1958 | break; |
1959 | } | |
1960 | } | |
eafe8130 TL |
1961 | if (!to_cancel->empty() && to_cancel->back() == pg) |
1962 | continue; | |
1963 | // okay, upmap is valid | |
1964 | // continue to check if it is still necessary | |
1965 | auto i = pg_upmap.find(pg); | |
a4b75251 TL |
1966 | if (i != pg_upmap.end()) { |
1967 | if (i->second == raw) { | |
1968 | ldout(cct, 10) << "removing redundant pg_upmap " << i->first << " " | |
1969 | << i->second << dendl; | |
1970 | to_cancel->push_back(pg); | |
1971 | continue; | |
1972 | } | |
1973 | if ((int)i->second.size() != get_pg_pool_size(pg)) { | |
1974 | ldout(cct, 10) << "removing pg_upmap " << i->first << " " | |
1975 | << i->second << " != pool size " << get_pg_pool_size(pg) | |
1976 | << dendl; | |
1977 | to_cancel->push_back(pg); | |
1978 | continue; | |
1979 | } | |
eafe8130 TL |
1980 | } |
1981 | auto j = pg_upmap_items.find(pg); | |
1982 | if (j != pg_upmap_items.end()) { | |
1983 | mempool::osdmap::vector<pair<int,int>> newmap; | |
1984 | for (auto& p : j->second) { | |
1985 | if (std::find(raw.begin(), raw.end(), p.first) == raw.end()) { | |
1986 | // cancel mapping if source osd does not exist anymore | |
1987 | continue; | |
1988 | } | |
1989 | if (p.second != CRUSH_ITEM_NONE && p.second < max_osd && | |
1990 | p.second >= 0 && osd_weight[p.second] == 0) { | |
1991 | // cancel mapping if target osd is out | |
1992 | continue; | |
1993 | } | |
1994 | newmap.push_back(p); | |
1995 | } | |
1996 | if (newmap.empty()) { | |
1997 | ldout(cct, 10) << " removing no-op pg_upmap_items " | |
1998 | << j->first << " " << j->second | |
1999 | << dendl; | |
2000 | to_cancel->push_back(pg); | |
2001 | } else if (newmap != j->second) { | |
2002 | ldout(cct, 10) << " simplifying partially no-op pg_upmap_items " | |
2003 | << j->first << " " << j->second | |
2004 | << " -> " << newmap | |
2005 | << dendl; | |
2006 | to_remap->insert({pg, newmap}); | |
2007 | any_change = true; | |
2008 | } | |
2009 | } | |
28e407b8 | 2010 | } |
494da23a TL |
2011 | any_change = any_change || !to_cancel->empty(); |
2012 | return any_change; | |
2013 | } | |
2014 | ||
2015 | void OSDMap::clean_pg_upmaps( | |
2016 | CephContext *cct, | |
2017 | Incremental *pending_inc, | |
2018 | const vector<pg_t>& to_cancel, | |
2019 | const map<pg_t, mempool::osdmap::vector<pair<int,int>>>& to_remap) const | |
2020 | { | |
28e407b8 | 2021 | for (auto &pg: to_cancel) { |
494da23a TL |
2022 | auto i = pending_inc->new_pg_upmap.find(pg); |
2023 | if (i != pending_inc->new_pg_upmap.end()) { | |
2024 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2025 | << "pg_upmap entry " | |
2026 | << i->first << "->" << i->second | |
2027 | << dendl; | |
2028 | pending_inc->new_pg_upmap.erase(i); | |
94b18763 | 2029 | } |
494da23a TL |
2030 | auto j = pg_upmap.find(pg); |
2031 | if (j != pg_upmap.end()) { | |
2032 | ldout(cct, 10) << __func__ << " cancel invalid pg_upmap entry " | |
2033 | << j->first << "->" << j->second | |
2034 | << dendl; | |
2035 | pending_inc->old_pg_upmap.insert(pg); | |
2036 | } | |
2037 | auto p = pending_inc->new_pg_upmap_items.find(pg); | |
2038 | if (p != pending_inc->new_pg_upmap_items.end()) { | |
2039 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2040 | << "pg_upmap_items entry " | |
2041 | << p->first << "->" << p->second | |
2042 | << dendl; | |
2043 | pending_inc->new_pg_upmap_items.erase(p); | |
2044 | } | |
2045 | auto q = pg_upmap_items.find(pg); | |
2046 | if (q != pg_upmap_items.end()) { | |
2047 | ldout(cct, 10) << __func__ << " cancel invalid " | |
2048 | << "pg_upmap_items entry " | |
2049 | << q->first << "->" << q->second | |
2050 | << dendl; | |
2051 | pending_inc->old_pg_upmap_items.insert(pg); | |
94b18763 FG |
2052 | } |
2053 | } | |
494da23a TL |
2054 | for (auto& i : to_remap) |
2055 | pending_inc->new_pg_upmap_items[i.first] = i.second; | |
2056 | } | |
2057 | ||
2058 | bool OSDMap::clean_pg_upmaps( | |
2059 | CephContext *cct, | |
2060 | Incremental *pending_inc) const | |
2061 | { | |
2062 | ldout(cct, 10) << __func__ << dendl; | |
2063 | vector<pg_t> to_check; | |
2064 | vector<pg_t> to_cancel; | |
2065 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> to_remap; | |
2066 | ||
2067 | get_upmap_pgs(&to_check); | |
2068 | auto any_change = check_pg_upmaps(cct, to_check, &to_cancel, &to_remap); | |
2069 | clean_pg_upmaps(cct, pending_inc, to_cancel, to_remap); | |
2070 | return any_change; | |
94b18763 FG |
2071 | } |
2072 | ||
7c673cae FG |
2073 | int OSDMap::apply_incremental(const Incremental &inc) |
2074 | { | |
f67539c2 | 2075 | new_blocklist_entries = false; |
7c673cae FG |
2076 | if (inc.epoch == 1) |
2077 | fsid = inc.fsid; | |
2078 | else if (inc.fsid != fsid) | |
2079 | return -EINVAL; | |
2080 | ||
11fdf7f2 | 2081 | ceph_assert(inc.epoch == epoch+1); |
7c673cae FG |
2082 | |
2083 | epoch++; | |
2084 | modified = inc.modified; | |
2085 | ||
2086 | // full map? | |
2087 | if (inc.fullmap.length()) { | |
9f95a23c | 2088 | ceph::buffer::list bl(inc.fullmap); |
7c673cae FG |
2089 | decode(bl); |
2090 | return 0; | |
2091 | } | |
2092 | ||
2093 | // nope, incremental. | |
31f18b77 | 2094 | if (inc.new_flags >= 0) { |
7c673cae | 2095 | flags = inc.new_flags; |
31f18b77 FG |
2096 | // the below is just to cover a newly-upgraded luminous mon |
2097 | // cluster that has to set require_jewel_osds or | |
2098 | // require_kraken_osds before the osds can be upgraded to | |
2099 | // luminous. | |
2100 | if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { | |
9f95a23c TL |
2101 | if (require_osd_release < ceph_release_t::kraken) { |
2102 | require_osd_release = ceph_release_t::kraken; | |
31f18b77 FG |
2103 | } |
2104 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
9f95a23c TL |
2105 | if (require_osd_release < ceph_release_t::jewel) { |
2106 | require_osd_release = ceph_release_t::jewel; | |
31f18b77 FG |
2107 | } |
2108 | } | |
2109 | } | |
7c673cae FG |
2110 | |
2111 | if (inc.new_max_osd >= 0) | |
2112 | set_max_osd(inc.new_max_osd); | |
2113 | ||
2114 | if (inc.new_pool_max != -1) | |
2115 | pool_max = inc.new_pool_max; | |
2116 | ||
2117 | for (const auto &pool : inc.new_pools) { | |
2118 | pools[pool.first] = pool.second; | |
2119 | pools[pool.first].last_change = epoch; | |
2120 | } | |
2121 | ||
11fdf7f2 TL |
2122 | new_removed_snaps = inc.new_removed_snaps; |
2123 | new_purged_snaps = inc.new_purged_snaps; | |
2124 | for (auto p = new_removed_snaps.begin(); | |
2125 | p != new_removed_snaps.end(); | |
2126 | ++p) { | |
2127 | removed_snaps_queue[p->first].union_of(p->second); | |
2128 | } | |
2129 | for (auto p = new_purged_snaps.begin(); | |
2130 | p != new_purged_snaps.end(); | |
2131 | ++p) { | |
2132 | auto q = removed_snaps_queue.find(p->first); | |
2133 | ceph_assert(q != removed_snaps_queue.end()); | |
2134 | q->second.subtract(p->second); | |
2135 | if (q->second.empty()) { | |
2136 | removed_snaps_queue.erase(q); | |
2137 | } | |
2138 | } | |
2139 | ||
2140 | if (inc.new_last_up_change != utime_t()) { | |
2141 | last_up_change = inc.new_last_up_change; | |
2142 | } | |
2143 | if (inc.new_last_in_change != utime_t()) { | |
2144 | last_in_change = inc.new_last_in_change; | |
2145 | } | |
2146 | ||
7c673cae FG |
2147 | for (const auto &pname : inc.new_pool_names) { |
2148 | auto pool_name_entry = pool_name.find(pname.first); | |
2149 | if (pool_name_entry != pool_name.end()) { | |
2150 | name_pool.erase(pool_name_entry->second); | |
2151 | pool_name_entry->second = pname.second; | |
2152 | } else { | |
2153 | pool_name[pname.first] = pname.second; | |
2154 | } | |
2155 | name_pool[pname.second] = pname.first; | |
2156 | } | |
2157 | ||
2158 | for (const auto &pool : inc.old_pools) { | |
2159 | pools.erase(pool); | |
2160 | name_pool.erase(pool_name[pool]); | |
2161 | pool_name.erase(pool); | |
2162 | } | |
2163 | ||
2164 | for (const auto &weight : inc.new_weight) { | |
2165 | set_weight(weight.first, weight.second); | |
2166 | ||
2167 | // if we are marking in, clear the AUTOOUT and NEW bits, and clear | |
2168 | // xinfo old_weight. | |
2169 | if (weight.second) { | |
2170 | osd_state[weight.first] &= ~(CEPH_OSD_AUTOOUT | CEPH_OSD_NEW); | |
2171 | osd_xinfo[weight.first].old_weight = 0; | |
2172 | } | |
2173 | } | |
2174 | ||
2175 | for (const auto &primary_affinity : inc.new_primary_affinity) { | |
2176 | set_primary_affinity(primary_affinity.first, primary_affinity.second); | |
2177 | } | |
2178 | ||
2179 | // erasure_code_profiles | |
2180 | for (const auto &profile : inc.old_erasure_code_profiles) | |
2181 | erasure_code_profiles.erase(profile); | |
2182 | ||
2183 | for (const auto &profile : inc.new_erasure_code_profiles) { | |
2184 | set_erasure_code_profile(profile.first, profile.second); | |
2185 | } | |
2186 | ||
2187 | // up/down | |
2188 | for (const auto &state : inc.new_state) { | |
2189 | const auto osd = state.first; | |
2190 | int s = state.second ? state.second : CEPH_OSD_UP; | |
2191 | if ((osd_state[osd] & CEPH_OSD_UP) && | |
2192 | (s & CEPH_OSD_UP)) { | |
2193 | osd_info[osd].down_at = epoch; | |
2194 | osd_xinfo[osd].down_stamp = modified; | |
2195 | } | |
2196 | if ((osd_state[osd] & CEPH_OSD_EXISTS) && | |
2197 | (s & CEPH_OSD_EXISTS)) { | |
2198 | // osd is destroyed; clear out anything interesting. | |
2199 | (*osd_uuid)[osd] = uuid_d(); | |
2200 | osd_info[osd] = osd_info_t(); | |
2201 | osd_xinfo[osd] = osd_xinfo_t(); | |
2202 | set_primary_affinity(osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
11fdf7f2 TL |
2203 | osd_addrs->client_addrs[osd].reset(new entity_addrvec_t()); |
2204 | osd_addrs->cluster_addrs[osd].reset(new entity_addrvec_t()); | |
2205 | osd_addrs->hb_front_addrs[osd].reset(new entity_addrvec_t()); | |
2206 | osd_addrs->hb_back_addrs[osd].reset(new entity_addrvec_t()); | |
7c673cae FG |
2207 | osd_state[osd] = 0; |
2208 | } else { | |
2209 | osd_state[osd] ^= s; | |
2210 | } | |
2211 | } | |
2212 | ||
2213 | for (const auto &client : inc.new_up_client) { | |
2214 | osd_state[client.first] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; | |
9f95a23c | 2215 | osd_state[client.first] &= ~CEPH_OSD_STOP; // if any |
11fdf7f2 TL |
2216 | osd_addrs->client_addrs[client.first].reset( |
2217 | new entity_addrvec_t(client.second)); | |
2218 | osd_addrs->hb_back_addrs[client.first].reset( | |
2219 | new entity_addrvec_t(inc.new_hb_back_up.find(client.first)->second)); | |
2220 | osd_addrs->hb_front_addrs[client.first].reset( | |
2221 | new entity_addrvec_t(inc.new_hb_front_up.find(client.first)->second)); | |
7c673cae FG |
2222 | |
2223 | osd_info[client.first].up_from = epoch; | |
2224 | } | |
2225 | ||
2226 | for (const auto &cluster : inc.new_up_cluster) | |
11fdf7f2 TL |
2227 | osd_addrs->cluster_addrs[cluster.first].reset( |
2228 | new entity_addrvec_t(cluster.second)); | |
7c673cae FG |
2229 | |
2230 | // info | |
2231 | for (const auto &thru : inc.new_up_thru) | |
2232 | osd_info[thru.first].up_thru = thru.second; | |
2233 | ||
2234 | for (const auto &interval : inc.new_last_clean_interval) { | |
2235 | osd_info[interval.first].last_clean_begin = interval.second.first; | |
2236 | osd_info[interval.first].last_clean_end = interval.second.second; | |
2237 | } | |
2238 | ||
2239 | for (const auto &lost : inc.new_lost) | |
2240 | osd_info[lost.first].lost_at = lost.second; | |
2241 | ||
2242 | // xinfo | |
2243 | for (const auto &xinfo : inc.new_xinfo) | |
2244 | osd_xinfo[xinfo.first] = xinfo.second; | |
2245 | ||
2246 | // uuid | |
2247 | for (const auto &uuid : inc.new_uuid) | |
2248 | (*osd_uuid)[uuid.first] = uuid.second; | |
2249 | ||
2250 | // pg rebuild | |
2251 | for (const auto &pg : inc.new_pg_temp) { | |
2252 | if (pg.second.empty()) | |
2253 | pg_temp->erase(pg.first); | |
2254 | else | |
31f18b77 FG |
2255 | pg_temp->set(pg.first, pg.second); |
2256 | } | |
2257 | if (!inc.new_pg_temp.empty()) { | |
2258 | // make sure pg_temp is efficiently stored | |
2259 | pg_temp->rebuild(); | |
7c673cae FG |
2260 | } |
2261 | ||
2262 | for (const auto &pg : inc.new_primary_temp) { | |
2263 | if (pg.second == -1) | |
2264 | primary_temp->erase(pg.first); | |
2265 | else | |
2266 | (*primary_temp)[pg.first] = pg.second; | |
2267 | } | |
2268 | ||
2269 | for (auto& p : inc.new_pg_upmap) { | |
2270 | pg_upmap[p.first] = p.second; | |
2271 | } | |
2272 | for (auto& pg : inc.old_pg_upmap) { | |
2273 | pg_upmap.erase(pg); | |
2274 | } | |
2275 | for (auto& p : inc.new_pg_upmap_items) { | |
2276 | pg_upmap_items[p.first] = p.second; | |
2277 | } | |
2278 | for (auto& pg : inc.old_pg_upmap_items) { | |
2279 | pg_upmap_items.erase(pg); | |
2280 | } | |
2281 | ||
f67539c2 TL |
2282 | // blocklist |
2283 | if (!inc.new_blocklist.empty()) { | |
2284 | blocklist.insert(inc.new_blocklist.begin(),inc.new_blocklist.end()); | |
2285 | new_blocklist_entries = true; | |
7c673cae | 2286 | } |
f67539c2 TL |
2287 | for (const auto &addr : inc.old_blocklist) |
2288 | blocklist.erase(addr); | |
7c673cae | 2289 | |
81eedcae TL |
2290 | for (auto& i : inc.new_crush_node_flags) { |
2291 | if (i.second) { | |
2292 | crush_node_flags[i.first] = i.second; | |
2293 | } else { | |
2294 | crush_node_flags.erase(i.first); | |
2295 | } | |
2296 | } | |
2297 | ||
2298 | for (auto& i : inc.new_device_class_flags) { | |
2299 | if (i.second) { | |
2300 | device_class_flags[i.first] = i.second; | |
2301 | } else { | |
2302 | device_class_flags.erase(i.first); | |
2303 | } | |
2304 | } | |
2305 | ||
7c673cae FG |
2306 | // cluster snapshot? |
2307 | if (inc.cluster_snapshot.length()) { | |
2308 | cluster_snapshot = inc.cluster_snapshot; | |
2309 | cluster_snapshot_epoch = inc.epoch; | |
2310 | } else { | |
2311 | cluster_snapshot.clear(); | |
2312 | cluster_snapshot_epoch = 0; | |
2313 | } | |
2314 | ||
2315 | if (inc.new_nearfull_ratio >= 0) { | |
2316 | nearfull_ratio = inc.new_nearfull_ratio; | |
2317 | } | |
2318 | if (inc.new_backfillfull_ratio >= 0) { | |
2319 | backfillfull_ratio = inc.new_backfillfull_ratio; | |
2320 | } | |
2321 | if (inc.new_full_ratio >= 0) { | |
2322 | full_ratio = inc.new_full_ratio; | |
2323 | } | |
9f95a23c | 2324 | if (inc.new_require_min_compat_client > ceph_release_t::unknown) { |
7c673cae FG |
2325 | require_min_compat_client = inc.new_require_min_compat_client; |
2326 | } | |
9f95a23c | 2327 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
31f18b77 | 2328 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2329 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 2330 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 2331 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2332 | } |
2333 | } | |
7c673cae | 2334 | |
9f95a23c | 2335 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
11fdf7f2 | 2336 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2337 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
2338 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
2339 | } | |
2340 | } | |
7c673cae FG |
2341 | // do new crush map last (after up/down stuff) |
2342 | if (inc.crush.length()) { | |
9f95a23c | 2343 | ceph::buffer::list bl(inc.crush); |
11fdf7f2 | 2344 | auto blp = bl.cbegin(); |
7c673cae FG |
2345 | crush.reset(new CrushWrapper); |
2346 | crush->decode(blp); | |
9f95a23c | 2347 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 FG |
2348 | // only increment if this is a luminous-encoded osdmap, lest |
2349 | // the mon's crush_version diverge from what the osds or others | |
2350 | // are decoding and applying on their end. if we won't encode | |
2351 | // it in the canonical version, don't change it. | |
2352 | ++crush_version; | |
2353 | } | |
81eedcae TL |
2354 | for (auto it = device_class_flags.begin(); |
2355 | it != device_class_flags.end();) { | |
2356 | const char* class_name = crush->get_class_name(it->first); | |
2357 | if (!class_name) // device class is gone | |
2358 | it = device_class_flags.erase(it); | |
2359 | else | |
2360 | it++; | |
2361 | } | |
7c673cae FG |
2362 | } |
2363 | ||
f67539c2 TL |
2364 | if (inc.change_stretch_mode) { |
2365 | stretch_mode_enabled = inc.stretch_mode_enabled; | |
2366 | stretch_bucket_count = inc.new_stretch_bucket_count; | |
2367 | degraded_stretch_mode = inc.new_degraded_stretch_mode; | |
2368 | recovering_stretch_mode = inc.new_recovering_stretch_mode; | |
2369 | stretch_mode_bucket = inc.new_stretch_mode_bucket; | |
2370 | } | |
2371 | ||
7c673cae FG |
2372 | calc_num_osds(); |
2373 | _calc_up_osd_features(); | |
2374 | return 0; | |
2375 | } | |
2376 | ||
2377 | // mapping | |
2378 | int OSDMap::map_to_pg( | |
2379 | int64_t poolid, | |
2380 | const string& name, | |
2381 | const string& key, | |
2382 | const string& nspace, | |
2383 | pg_t *pg) const | |
2384 | { | |
2385 | // calculate ps (placement seed) | |
2386 | const pg_pool_t *pool = get_pg_pool(poolid); | |
2387 | if (!pool) | |
2388 | return -ENOENT; | |
2389 | ps_t ps; | |
2390 | if (!key.empty()) | |
2391 | ps = pool->hash_key(key, nspace); | |
2392 | else | |
2393 | ps = pool->hash_key(name, nspace); | |
2394 | *pg = pg_t(ps, poolid); | |
2395 | return 0; | |
2396 | } | |
2397 | ||
2398 | int OSDMap::object_locator_to_pg( | |
2399 | const object_t& oid, const object_locator_t& loc, pg_t &pg) const | |
2400 | { | |
2401 | if (loc.hash >= 0) { | |
2402 | if (!get_pg_pool(loc.get_pool())) { | |
2403 | return -ENOENT; | |
2404 | } | |
2405 | pg = pg_t(loc.hash, loc.get_pool()); | |
2406 | return 0; | |
2407 | } | |
2408 | return map_to_pg(loc.get_pool(), oid.name, loc.key, loc.nspace, &pg); | |
2409 | } | |
2410 | ||
2411 | ceph_object_layout OSDMap::make_object_layout( | |
2412 | object_t oid, int pg_pool, string nspace) const | |
2413 | { | |
2414 | object_locator_t loc(pg_pool, nspace); | |
2415 | ||
2416 | ceph_object_layout ol; | |
2417 | pg_t pgid = object_locator_to_pg(oid, loc); | |
2418 | ol.ol_pgid = pgid.get_old_pg().v; | |
2419 | ol.ol_stripe_unit = 0; | |
2420 | return ol; | |
2421 | } | |
2422 | ||
2423 | void OSDMap::_remove_nonexistent_osds(const pg_pool_t& pool, | |
2424 | vector<int>& osds) const | |
2425 | { | |
2426 | if (pool.can_shift_osds()) { | |
2427 | unsigned removed = 0; | |
2428 | for (unsigned i = 0; i < osds.size(); i++) { | |
2429 | if (!exists(osds[i])) { | |
2430 | removed++; | |
2431 | continue; | |
2432 | } | |
2433 | if (removed) { | |
2434 | osds[i - removed] = osds[i]; | |
2435 | } | |
2436 | } | |
2437 | if (removed) | |
2438 | osds.resize(osds.size() - removed); | |
2439 | } else { | |
2440 | for (auto& osd : osds) { | |
2441 | if (!exists(osd)) | |
2442 | osd = CRUSH_ITEM_NONE; | |
2443 | } | |
2444 | } | |
2445 | } | |
2446 | ||
31f18b77 | 2447 | void OSDMap::_pg_to_raw_osds( |
7c673cae FG |
2448 | const pg_pool_t& pool, pg_t pg, |
2449 | vector<int> *osds, | |
2450 | ps_t *ppps) const | |
2451 | { | |
2452 | // map to osds[] | |
2453 | ps_t pps = pool.raw_pg_to_pps(pg); // placement ps | |
2454 | unsigned size = pool.get_size(); | |
2455 | ||
2456 | // what crush rule? | |
20effc67 | 2457 | int ruleno = pool.get_crush_rule(); |
7c673cae FG |
2458 | if (ruleno >= 0) |
2459 | crush->do_rule(ruleno, pps, *osds, size, osd_weight, pg.pool()); | |
2460 | ||
2461 | _remove_nonexistent_osds(pool, *osds); | |
2462 | ||
2463 | if (ppps) | |
2464 | *ppps = pps; | |
7c673cae FG |
2465 | } |
2466 | ||
2467 | int OSDMap::_pick_primary(const vector<int>& osds) const | |
2468 | { | |
2469 | for (auto osd : osds) { | |
2470 | if (osd != CRUSH_ITEM_NONE) { | |
2471 | return osd; | |
2472 | } | |
2473 | } | |
2474 | return -1; | |
2475 | } | |
2476 | ||
224ce89b | 2477 | void OSDMap::_apply_upmap(const pg_pool_t& pi, pg_t raw_pg, vector<int> *raw) const |
7c673cae FG |
2478 | { |
2479 | pg_t pg = pi.raw_pg_to_pg(raw_pg); | |
2480 | auto p = pg_upmap.find(pg); | |
2481 | if (p != pg_upmap.end()) { | |
2482 | // make sure targets aren't marked out | |
2483 | for (auto osd : p->second) { | |
91327a77 AA |
2484 | if (osd != CRUSH_ITEM_NONE && osd < max_osd && osd >= 0 && |
2485 | osd_weight[osd] == 0) { | |
7c673cae FG |
2486 | // reject/ignore the explicit mapping |
2487 | return; | |
2488 | } | |
2489 | } | |
2490 | *raw = vector<int>(p->second.begin(), p->second.end()); | |
224ce89b | 2491 | // continue to check and apply pg_upmap_items if any |
7c673cae FG |
2492 | } |
2493 | ||
2494 | auto q = pg_upmap_items.find(pg); | |
2495 | if (q != pg_upmap_items.end()) { | |
181888fb FG |
2496 | // NOTE: this approach does not allow a bidirectional swap, |
2497 | // e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1]. | |
2498 | for (auto& r : q->second) { | |
2499 | // make sure the replacement value doesn't already appear | |
2500 | bool exists = false; | |
2501 | ssize_t pos = -1; | |
2502 | for (unsigned i = 0; i < raw->size(); ++i) { | |
2503 | int osd = (*raw)[i]; | |
2504 | if (osd == r.second) { | |
2505 | exists = true; | |
2506 | break; | |
2507 | } | |
2508 | // ignore mapping if target is marked out (or invalid osd id) | |
2509 | if (osd == r.first && | |
2510 | pos < 0 && | |
2511 | !(r.second != CRUSH_ITEM_NONE && r.second < max_osd && | |
91327a77 | 2512 | r.second >= 0 && osd_weight[r.second] == 0)) { |
181888fb FG |
2513 | pos = i; |
2514 | } | |
2515 | } | |
2516 | if (!exists && pos >= 0) { | |
2517 | (*raw)[pos] = r.second; | |
7c673cae FG |
2518 | } |
2519 | } | |
2520 | } | |
2521 | } | |
2522 | ||
2523 | // pg -> (up osd list) | |
2524 | void OSDMap::_raw_to_up_osds(const pg_pool_t& pool, const vector<int>& raw, | |
2525 | vector<int> *up) const | |
2526 | { | |
2527 | if (pool.can_shift_osds()) { | |
2528 | // shift left | |
2529 | up->clear(); | |
2530 | up->reserve(raw.size()); | |
2531 | for (unsigned i=0; i<raw.size(); i++) { | |
2532 | if (!exists(raw[i]) || is_down(raw[i])) | |
2533 | continue; | |
2534 | up->push_back(raw[i]); | |
2535 | } | |
2536 | } else { | |
2537 | // set down/dne devices to NONE | |
2538 | up->resize(raw.size()); | |
2539 | for (int i = raw.size() - 1; i >= 0; --i) { | |
2540 | if (!exists(raw[i]) || is_down(raw[i])) { | |
2541 | (*up)[i] = CRUSH_ITEM_NONE; | |
2542 | } else { | |
2543 | (*up)[i] = raw[i]; | |
2544 | } | |
2545 | } | |
2546 | } | |
2547 | } | |
2548 | ||
2549 | void OSDMap::_apply_primary_affinity(ps_t seed, | |
2550 | const pg_pool_t& pool, | |
2551 | vector<int> *osds, | |
2552 | int *primary) const | |
2553 | { | |
2554 | // do we have any non-default primary_affinity values for these osds? | |
2555 | if (!osd_primary_affinity) | |
2556 | return; | |
2557 | ||
2558 | bool any = false; | |
2559 | for (const auto osd : *osds) { | |
2560 | if (osd != CRUSH_ITEM_NONE && | |
2561 | (*osd_primary_affinity)[osd] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
2562 | any = true; | |
2563 | break; | |
2564 | } | |
2565 | } | |
2566 | if (!any) | |
2567 | return; | |
2568 | ||
2569 | // pick the primary. feed both the seed (for the pg) and the osd | |
2570 | // into the hash/rng so that a proportional fraction of an osd's pgs | |
2571 | // get rejected as primary. | |
2572 | int pos = -1; | |
2573 | for (unsigned i = 0; i < osds->size(); ++i) { | |
2574 | int o = (*osds)[i]; | |
2575 | if (o == CRUSH_ITEM_NONE) | |
2576 | continue; | |
2577 | unsigned a = (*osd_primary_affinity)[o]; | |
2578 | if (a < CEPH_OSD_MAX_PRIMARY_AFFINITY && | |
2579 | (crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
2580 | seed, o) >> 16) >= a) { | |
2581 | // we chose not to use this primary. note it anyway as a | |
2582 | // fallback in case we don't pick anyone else, but keep looking. | |
2583 | if (pos < 0) | |
2584 | pos = i; | |
2585 | } else { | |
2586 | pos = i; | |
2587 | break; | |
2588 | } | |
2589 | } | |
2590 | if (pos < 0) | |
2591 | return; | |
2592 | ||
2593 | *primary = (*osds)[pos]; | |
2594 | ||
2595 | if (pool.can_shift_osds() && pos > 0) { | |
2596 | // move the new primary to the front. | |
2597 | for (int i = pos; i > 0; --i) { | |
2598 | (*osds)[i] = (*osds)[i-1]; | |
2599 | } | |
2600 | (*osds)[0] = *primary; | |
2601 | } | |
2602 | } | |
2603 | ||
2604 | void OSDMap::_get_temp_osds(const pg_pool_t& pool, pg_t pg, | |
2605 | vector<int> *temp_pg, int *temp_primary) const | |
2606 | { | |
2607 | pg = pool.raw_pg_to_pg(pg); | |
2608 | const auto p = pg_temp->find(pg); | |
2609 | temp_pg->clear(); | |
2610 | if (p != pg_temp->end()) { | |
2611 | for (unsigned i=0; i<p->second.size(); i++) { | |
2612 | if (!exists(p->second[i]) || is_down(p->second[i])) { | |
2613 | if (pool.can_shift_osds()) { | |
2614 | continue; | |
2615 | } else { | |
2616 | temp_pg->push_back(CRUSH_ITEM_NONE); | |
2617 | } | |
2618 | } else { | |
2619 | temp_pg->push_back(p->second[i]); | |
2620 | } | |
2621 | } | |
2622 | } | |
2623 | const auto &pp = primary_temp->find(pg); | |
2624 | *temp_primary = -1; | |
2625 | if (pp != primary_temp->end()) { | |
2626 | *temp_primary = pp->second; | |
2627 | } else if (!temp_pg->empty()) { // apply pg_temp's primary | |
2628 | for (unsigned i = 0; i < temp_pg->size(); ++i) { | |
2629 | if ((*temp_pg)[i] != CRUSH_ITEM_NONE) { | |
2630 | *temp_primary = (*temp_pg)[i]; | |
2631 | break; | |
2632 | } | |
2633 | } | |
2634 | } | |
2635 | } | |
2636 | ||
31f18b77 | 2637 | void OSDMap::pg_to_raw_osds(pg_t pg, vector<int> *raw, int *primary) const |
7c673cae | 2638 | { |
7c673cae | 2639 | const pg_pool_t *pool = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2640 | if (!pool) { |
2641 | *primary = -1; | |
2642 | raw->clear(); | |
31f18b77 | 2643 | return; |
11fdf7f2 | 2644 | } |
31f18b77 | 2645 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
11fdf7f2 | 2646 | *primary = _pick_primary(*raw); |
7c673cae FG |
2647 | } |
2648 | ||
494da23a TL |
2649 | void OSDMap::pg_to_raw_upmap(pg_t pg, vector<int>*raw, |
2650 | vector<int> *raw_upmap) const | |
a8e16298 TL |
2651 | { |
2652 | auto pool = get_pg_pool(pg.pool()); | |
2653 | if (!pool) { | |
2654 | raw_upmap->clear(); | |
2655 | return; | |
2656 | } | |
494da23a TL |
2657 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
2658 | *raw_upmap = *raw; | |
a8e16298 TL |
2659 | _apply_upmap(*pool, pg, raw_upmap); |
2660 | } | |
2661 | ||
7c673cae FG |
2662 | void OSDMap::pg_to_raw_up(pg_t pg, vector<int> *up, int *primary) const |
2663 | { | |
2664 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2665 | if (!pool) { | |
11fdf7f2 TL |
2666 | *primary = -1; |
2667 | up->clear(); | |
7c673cae FG |
2668 | return; |
2669 | } | |
2670 | vector<int> raw; | |
2671 | ps_t pps; | |
2672 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2673 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2674 | _raw_to_up_osds(*pool, raw, up); |
2675 | *primary = _pick_primary(raw); | |
2676 | _apply_primary_affinity(pps, *pool, up, primary); | |
2677 | } | |
31f18b77 | 2678 | |
7c673cae FG |
2679 | void OSDMap::_pg_to_up_acting_osds( |
2680 | const pg_t& pg, vector<int> *up, int *up_primary, | |
2681 | vector<int> *acting, int *acting_primary, | |
2682 | bool raw_pg_to_pg) const | |
2683 | { | |
2684 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2685 | if (!pool || | |
2686 | (!raw_pg_to_pg && pg.ps() >= pool->get_pg_num())) { | |
2687 | if (up) | |
2688 | up->clear(); | |
2689 | if (up_primary) | |
2690 | *up_primary = -1; | |
2691 | if (acting) | |
2692 | acting->clear(); | |
2693 | if (acting_primary) | |
2694 | *acting_primary = -1; | |
2695 | return; | |
2696 | } | |
2697 | vector<int> raw; | |
2698 | vector<int> _up; | |
2699 | vector<int> _acting; | |
2700 | int _up_primary; | |
2701 | int _acting_primary; | |
2702 | ps_t pps; | |
2703 | _get_temp_osds(*pool, pg, &_acting, &_acting_primary); | |
2704 | if (_acting.empty() || up || up_primary) { | |
2705 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2706 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2707 | _raw_to_up_osds(*pool, raw, &_up); |
2708 | _up_primary = _pick_primary(_up); | |
2709 | _apply_primary_affinity(pps, *pool, &_up, &_up_primary); | |
2710 | if (_acting.empty()) { | |
2711 | _acting = _up; | |
2712 | if (_acting_primary == -1) { | |
2713 | _acting_primary = _up_primary; | |
2714 | } | |
2715 | } | |
2716 | ||
2717 | if (up) | |
2718 | up->swap(_up); | |
2719 | if (up_primary) | |
2720 | *up_primary = _up_primary; | |
2721 | } | |
2722 | ||
2723 | if (acting) | |
2724 | acting->swap(_acting); | |
2725 | if (acting_primary) | |
2726 | *acting_primary = _acting_primary; | |
2727 | } | |
2728 | ||
9f95a23c | 2729 | int OSDMap::calc_pg_role_broken(int osd, const vector<int>& acting, int nrep) |
7c673cae | 2730 | { |
9f95a23c TL |
2731 | // This implementation is broken for EC PGs since the osd may appear |
2732 | // multiple times in the acting set. See | |
2733 | // https://tracker.ceph.com/issues/43213 | |
7c673cae FG |
2734 | if (!nrep) |
2735 | nrep = acting.size(); | |
2736 | for (int i=0; i<nrep; i++) | |
2737 | if (acting[i] == osd) | |
2738 | return i; | |
2739 | return -1; | |
2740 | } | |
2741 | ||
9f95a23c | 2742 | int OSDMap::calc_pg_role(pg_shard_t who, const vector<int>& acting) |
7c673cae | 2743 | { |
9f95a23c TL |
2744 | int nrep = acting.size(); |
2745 | if (who.shard == shard_id_t::NO_SHARD) { | |
2746 | for (int i=0; i<nrep; i++) { | |
2747 | if (acting[i] == who.osd) { | |
2748 | return i; | |
2749 | } | |
2750 | } | |
2751 | } else { | |
2752 | if (who.shard < nrep && acting[who.shard] == who.osd) { | |
2753 | return who.shard; | |
2754 | } | |
2755 | } | |
2756 | return -1; | |
7c673cae FG |
2757 | } |
2758 | ||
9f95a23c | 2759 | bool OSDMap::primary_changed_broken( |
7c673cae FG |
2760 | int oldprimary, |
2761 | const vector<int> &oldacting, | |
2762 | int newprimary, | |
2763 | const vector<int> &newacting) | |
2764 | { | |
2765 | if (oldacting.empty() && newacting.empty()) | |
2766 | return false; // both still empty | |
2767 | if (oldacting.empty() ^ newacting.empty()) | |
2768 | return true; // was empty, now not, or vice versa | |
2769 | if (oldprimary != newprimary) | |
2770 | return true; // primary changed | |
9f95a23c TL |
2771 | if (calc_pg_role_broken(oldprimary, oldacting) != |
2772 | calc_pg_role_broken(newprimary, newacting)) | |
7c673cae FG |
2773 | return true; |
2774 | return false; // same primary (tho replicas may have changed) | |
2775 | } | |
2776 | ||
28e407b8 AA |
2777 | uint64_t OSDMap::get_encoding_features() const |
2778 | { | |
2779 | uint64_t f = SIGNIFICANT_FEATURES; | |
9f95a23c TL |
2780 | if (require_osd_release < ceph_release_t::octopus) { |
2781 | f &= ~CEPH_FEATURE_SERVER_OCTOPUS; | |
2782 | } | |
2783 | if (require_osd_release < ceph_release_t::nautilus) { | |
11fdf7f2 TL |
2784 | f &= ~CEPH_FEATURE_SERVER_NAUTILUS; |
2785 | } | |
9f95a23c | 2786 | if (require_osd_release < ceph_release_t::mimic) { |
11fdf7f2 TL |
2787 | f &= ~CEPH_FEATURE_SERVER_MIMIC; |
2788 | } | |
9f95a23c | 2789 | if (require_osd_release < ceph_release_t::luminous) { |
28e407b8 AA |
2790 | f &= ~(CEPH_FEATURE_SERVER_LUMINOUS | |
2791 | CEPH_FEATURE_CRUSH_CHOOSE_ARGS); | |
2792 | } | |
9f95a23c | 2793 | if (require_osd_release < ceph_release_t::kraken) { |
28e407b8 | 2794 | f &= ~(CEPH_FEATURE_SERVER_KRAKEN | |
1adf2230 | 2795 | CEPH_FEATURE_MSG_ADDR2); |
28e407b8 | 2796 | } |
9f95a23c | 2797 | if (require_osd_release < ceph_release_t::jewel) { |
28e407b8 | 2798 | f &= ~(CEPH_FEATURE_SERVER_JEWEL | |
1adf2230 AA |
2799 | CEPH_FEATURE_NEW_OSDOP_ENCODING | |
2800 | CEPH_FEATURE_CRUSH_TUNABLES5); | |
28e407b8 AA |
2801 | } |
2802 | return f; | |
2803 | } | |
7c673cae FG |
2804 | |
2805 | // serialize, unserialize | |
9f95a23c | 2806 | void OSDMap::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 2807 | { |
11fdf7f2 | 2808 | using ceph::encode; |
7c673cae | 2809 | __u16 v = 5; |
11fdf7f2 | 2810 | encode(v, bl); |
7c673cae FG |
2811 | |
2812 | // base | |
11fdf7f2 TL |
2813 | encode(fsid, bl); |
2814 | encode(epoch, bl); | |
2815 | encode(created, bl); | |
2816 | encode(modified, bl); | |
7c673cae | 2817 | |
11fdf7f2 | 2818 | // for encode(pools, bl); |
7c673cae | 2819 | __u32 n = pools.size(); |
11fdf7f2 | 2820 | encode(n, bl); |
7c673cae FG |
2821 | |
2822 | for (const auto &pool : pools) { | |
2823 | n = pool.first; | |
11fdf7f2 TL |
2824 | encode(n, bl); |
2825 | encode(pool.second, bl, 0); | |
7c673cae | 2826 | } |
11fdf7f2 | 2827 | // for encode(pool_name, bl); |
7c673cae | 2828 | n = pool_name.size(); |
11fdf7f2 | 2829 | encode(n, bl); |
7c673cae FG |
2830 | for (const auto &pname : pool_name) { |
2831 | n = pname.first; | |
11fdf7f2 TL |
2832 | encode(n, bl); |
2833 | encode(pname.second, bl); | |
7c673cae | 2834 | } |
11fdf7f2 | 2835 | // for encode(pool_max, bl); |
7c673cae | 2836 | n = pool_max; |
11fdf7f2 | 2837 | encode(n, bl); |
7c673cae | 2838 | |
11fdf7f2 | 2839 | encode(flags, bl); |
7c673cae | 2840 | |
11fdf7f2 | 2841 | encode(max_osd, bl); |
31f18b77 FG |
2842 | { |
2843 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2844 | encode(n, bl); |
31f18b77 | 2845 | for (auto s : osd_state) { |
11fdf7f2 | 2846 | encode((uint8_t)s, bl); |
31f18b77 FG |
2847 | } |
2848 | } | |
11fdf7f2 TL |
2849 | encode(osd_weight, bl); |
2850 | encode(osd_addrs->client_addrs, bl, 0); | |
7c673cae | 2851 | |
11fdf7f2 | 2852 | // for encode(pg_temp, bl); |
7c673cae | 2853 | n = pg_temp->size(); |
11fdf7f2 | 2854 | encode(n, bl); |
f67539c2 | 2855 | for (const auto& pg : *pg_temp) { |
7c673cae | 2856 | old_pg_t opg = pg.first.get_old_pg(); |
11fdf7f2 TL |
2857 | encode(opg, bl); |
2858 | encode(pg.second, bl); | |
7c673cae FG |
2859 | } |
2860 | ||
2861 | // crush | |
9f95a23c | 2862 | ceph::buffer::list cbl; |
7c673cae | 2863 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 2864 | encode(cbl, bl); |
7c673cae FG |
2865 | } |
2866 | ||
9f95a23c | 2867 | void OSDMap::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 2868 | { |
11fdf7f2 | 2869 | using ceph::encode; |
7c673cae FG |
2870 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
2871 | encode_client_old(bl); | |
2872 | return; | |
2873 | } | |
2874 | ||
2875 | __u16 v = 6; | |
11fdf7f2 | 2876 | encode(v, bl); |
7c673cae FG |
2877 | |
2878 | // base | |
11fdf7f2 TL |
2879 | encode(fsid, bl); |
2880 | encode(epoch, bl); | |
2881 | encode(created, bl); | |
2882 | encode(modified, bl); | |
7c673cae | 2883 | |
11fdf7f2 TL |
2884 | encode(pools, bl, features); |
2885 | encode(pool_name, bl); | |
2886 | encode(pool_max, bl); | |
7c673cae | 2887 | |
11fdf7f2 | 2888 | encode(flags, bl); |
7c673cae | 2889 | |
11fdf7f2 | 2890 | encode(max_osd, bl); |
31f18b77 FG |
2891 | { |
2892 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2893 | encode(n, bl); |
31f18b77 | 2894 | for (auto s : osd_state) { |
11fdf7f2 | 2895 | encode((uint8_t)s, bl); |
31f18b77 FG |
2896 | } |
2897 | } | |
11fdf7f2 TL |
2898 | encode(osd_weight, bl); |
2899 | encode(osd_addrs->client_addrs, bl, features); | |
7c673cae | 2900 | |
11fdf7f2 | 2901 | encode(*pg_temp, bl); |
7c673cae FG |
2902 | |
2903 | // crush | |
9f95a23c | 2904 | ceph::buffer::list cbl; |
7c673cae | 2905 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 2906 | encode(cbl, bl); |
7c673cae FG |
2907 | |
2908 | // extended | |
2909 | __u16 ev = 10; | |
11fdf7f2 TL |
2910 | encode(ev, bl); |
2911 | encode(osd_addrs->hb_back_addrs, bl, features); | |
2912 | encode(osd_info, bl); | |
f67539c2 | 2913 | encode(blocklist, bl, features); |
11fdf7f2 TL |
2914 | encode(osd_addrs->cluster_addrs, bl, features); |
2915 | encode(cluster_snapshot_epoch, bl); | |
2916 | encode(cluster_snapshot, bl); | |
2917 | encode(*osd_uuid, bl); | |
9f95a23c | 2918 | encode(osd_xinfo, bl, features); |
11fdf7f2 | 2919 | encode(osd_addrs->hb_front_addrs, bl, features); |
7c673cae FG |
2920 | } |
2921 | ||
11fdf7f2 TL |
2922 | /* for a description of osdmap versions, and when they were introduced, please |
2923 | * refer to | |
2924 | * doc/dev/osd_internals/osdmap_versions.txt | |
2925 | */ | |
9f95a23c | 2926 | void OSDMap::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 2927 | { |
11fdf7f2 | 2928 | using ceph::encode; |
7c673cae FG |
2929 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
2930 | encode_classic(bl, features); | |
2931 | return; | |
2932 | } | |
2933 | ||
2934 | // only a select set of callers should *ever* be encoding new | |
2935 | // OSDMaps. others should be passing around the canonical encoded | |
2936 | // buffers from on high. select out those callers by passing in an | |
2937 | // "impossible" feature bit. | |
11fdf7f2 | 2938 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
2939 | features &= ~CEPH_FEATURE_RESERVED; |
2940 | ||
2941 | size_t start_offset = bl.length(); | |
2942 | size_t tail_offset; | |
11fdf7f2 | 2943 | size_t crc_offset; |
9f95a23c | 2944 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
2945 | |
2946 | // meta-encoding: how we include client-used and osd-specific data | |
2947 | ENCODE_START(8, 7, bl); | |
2948 | ||
2949 | { | |
28e407b8 AA |
2950 | // NOTE: any new encoding dependencies must be reflected by |
2951 | // SIGNIFICANT_FEATURES | |
11fdf7f2 | 2952 | uint8_t v = 9; |
31f18b77 | 2953 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
7c673cae | 2954 | v = 3; |
11fdf7f2 TL |
2955 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
2956 | v = 6; | |
2957 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
2958 | v = 7; | |
7c673cae FG |
2959 | } |
2960 | ENCODE_START(v, 1, bl); // client-usable data | |
2961 | // base | |
11fdf7f2 TL |
2962 | encode(fsid, bl); |
2963 | encode(epoch, bl); | |
2964 | encode(created, bl); | |
2965 | encode(modified, bl); | |
7c673cae | 2966 | |
11fdf7f2 TL |
2967 | encode(pools, bl, features); |
2968 | encode(pool_name, bl); | |
2969 | encode(pool_max, bl); | |
7c673cae | 2970 | |
31f18b77 FG |
2971 | if (v < 4) { |
2972 | decltype(flags) f = flags; | |
9f95a23c | 2973 | if (require_osd_release >= ceph_release_t::luminous) |
c07f9fc5 | 2974 | f |= CEPH_OSDMAP_REQUIRE_LUMINOUS | CEPH_OSDMAP_RECOVERY_DELETES; |
9f95a23c | 2975 | else if (require_osd_release == ceph_release_t::kraken) |
31f18b77 | 2976 | f |= CEPH_OSDMAP_REQUIRE_KRAKEN; |
9f95a23c | 2977 | else if (require_osd_release == ceph_release_t::jewel) |
31f18b77 | 2978 | f |= CEPH_OSDMAP_REQUIRE_JEWEL; |
11fdf7f2 | 2979 | encode(f, bl); |
31f18b77 | 2980 | } else { |
11fdf7f2 | 2981 | encode(flags, bl); |
31f18b77 | 2982 | } |
7c673cae | 2983 | |
11fdf7f2 | 2984 | encode(max_osd, bl); |
31f18b77 | 2985 | if (v >= 5) { |
11fdf7f2 | 2986 | encode(osd_state, bl); |
31f18b77 FG |
2987 | } else { |
2988 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2989 | encode(n, bl); |
31f18b77 | 2990 | for (auto s : osd_state) { |
11fdf7f2 | 2991 | encode((uint8_t)s, bl); |
31f18b77 FG |
2992 | } |
2993 | } | |
11fdf7f2 TL |
2994 | encode(osd_weight, bl); |
2995 | if (v >= 8) { | |
2996 | encode(osd_addrs->client_addrs, bl, features); | |
2997 | } else { | |
2998 | encode_addrvec_pvec_as_addr(osd_addrs->client_addrs, bl, features); | |
2999 | } | |
7c673cae | 3000 | |
11fdf7f2 TL |
3001 | encode(*pg_temp, bl); |
3002 | encode(*primary_temp, bl); | |
7c673cae | 3003 | if (osd_primary_affinity) { |
11fdf7f2 | 3004 | encode(*osd_primary_affinity, bl); |
7c673cae FG |
3005 | } else { |
3006 | vector<__u32> v; | |
11fdf7f2 | 3007 | encode(v, bl); |
7c673cae FG |
3008 | } |
3009 | ||
3010 | // crush | |
9f95a23c | 3011 | ceph::buffer::list cbl; |
7c673cae | 3012 | crush->encode(cbl, features); |
11fdf7f2 TL |
3013 | encode(cbl, bl); |
3014 | encode(erasure_code_profiles, bl); | |
7c673cae FG |
3015 | |
3016 | if (v >= 4) { | |
11fdf7f2 TL |
3017 | encode(pg_upmap, bl); |
3018 | encode(pg_upmap_items, bl); | |
7c673cae | 3019 | } else { |
11fdf7f2 TL |
3020 | ceph_assert(pg_upmap.empty()); |
3021 | ceph_assert(pg_upmap_items.empty()); | |
7c673cae | 3022 | } |
31f18b77 | 3023 | if (v >= 6) { |
11fdf7f2 TL |
3024 | encode(crush_version, bl); |
3025 | } | |
3026 | if (v >= 7) { | |
3027 | encode(new_removed_snaps, bl); | |
3028 | encode(new_purged_snaps, bl); | |
3029 | } | |
3030 | if (v >= 9) { | |
3031 | encode(last_up_change, bl); | |
3032 | encode(last_in_change, bl); | |
31f18b77 | 3033 | } |
7c673cae FG |
3034 | ENCODE_FINISH(bl); // client-usable data |
3035 | } | |
3036 | ||
3037 | { | |
28e407b8 AA |
3038 | // NOTE: any new encoding dependencies must be reflected by |
3039 | // SIGNIFICANT_FEATURES | |
f67539c2 | 3040 | uint8_t target_v = 9; // when bumping this, be aware of stretch_mode target_v 10! |
7c673cae FG |
3041 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
3042 | target_v = 1; | |
11fdf7f2 TL |
3043 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
3044 | target_v = 5; | |
3045 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
3046 | target_v = 6; | |
7c673cae | 3047 | } |
f67539c2 TL |
3048 | if (stretch_mode_enabled) { |
3049 | target_v = std::max((uint8_t)10, target_v); | |
3050 | } | |
7c673cae | 3051 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
3052 | if (target_v < 7) { |
3053 | encode_addrvec_pvec_as_addr(osd_addrs->hb_back_addrs, bl, features); | |
3054 | } else { | |
3055 | encode(osd_addrs->hb_back_addrs, bl, features); | |
3056 | } | |
3057 | encode(osd_info, bl); | |
7c673cae FG |
3058 | { |
3059 | // put this in a sorted, ordered map<> so that we encode in a | |
3060 | // deterministic order. | |
f67539c2 TL |
3061 | map<entity_addr_t,utime_t> blocklist_map; |
3062 | for (const auto &addr : blocklist) | |
3063 | blocklist_map.insert(make_pair(addr.first, addr.second)); | |
3064 | encode(blocklist_map, bl, features); | |
11fdf7f2 TL |
3065 | } |
3066 | if (target_v < 7) { | |
3067 | encode_addrvec_pvec_as_addr(osd_addrs->cluster_addrs, bl, features); | |
3068 | } else { | |
3069 | encode(osd_addrs->cluster_addrs, bl, features); | |
3070 | } | |
3071 | encode(cluster_snapshot_epoch, bl); | |
3072 | encode(cluster_snapshot, bl); | |
3073 | encode(*osd_uuid, bl); | |
9f95a23c | 3074 | encode(osd_xinfo, bl, features); |
11fdf7f2 TL |
3075 | if (target_v < 7) { |
3076 | encode_addrvec_pvec_as_addr(osd_addrs->hb_front_addrs, bl, features); | |
3077 | } else { | |
3078 | encode(osd_addrs->hb_front_addrs, bl, features); | |
3079 | } | |
7c673cae | 3080 | if (target_v >= 2) { |
11fdf7f2 TL |
3081 | encode(nearfull_ratio, bl); |
3082 | encode(full_ratio, bl); | |
3083 | encode(backfillfull_ratio, bl); | |
31f18b77 FG |
3084 | } |
3085 | // 4 was string-based new_require_min_compat_client | |
3086 | if (target_v >= 5) { | |
11fdf7f2 TL |
3087 | encode(require_min_compat_client, bl); |
3088 | encode(require_osd_release, bl); | |
3089 | } | |
3090 | if (target_v >= 6) { | |
3091 | encode(removed_snaps_queue, bl); | |
7c673cae | 3092 | } |
81eedcae TL |
3093 | if (target_v >= 8) { |
3094 | encode(crush_node_flags, bl); | |
3095 | } | |
3096 | if (target_v >= 9) { | |
3097 | encode(device_class_flags, bl); | |
3098 | } | |
f67539c2 TL |
3099 | if (target_v >= 10) { |
3100 | encode(stretch_mode_enabled, bl); | |
3101 | encode(stretch_bucket_count, bl); | |
3102 | encode(degraded_stretch_mode, bl); | |
3103 | encode(recovering_stretch_mode, bl); | |
3104 | encode(stretch_mode_bucket, bl); | |
3105 | } | |
7c673cae FG |
3106 | ENCODE_FINISH(bl); // osd-only data |
3107 | } | |
3108 | ||
11fdf7f2 TL |
3109 | crc_offset = bl.length(); |
3110 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
3111 | tail_offset = bl.length(); |
3112 | ||
3113 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
3114 | ||
3115 | // fill in crc | |
9f95a23c | 3116 | ceph::buffer::list front; |
11fdf7f2 | 3117 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae FG |
3118 | crc = front.crc32c(-1); |
3119 | if (tail_offset < bl.length()) { | |
9f95a23c | 3120 | ceph::buffer::list tail; |
7c673cae FG |
3121 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
3122 | crc = tail.crc32c(crc); | |
3123 | } | |
3124 | ceph_le32 crc_le; | |
3125 | crc_le = crc; | |
11fdf7f2 | 3126 | crc_filler->copy_in(4, (char*)&crc_le); |
7c673cae FG |
3127 | crc_defined = true; |
3128 | } | |
3129 | ||
11fdf7f2 TL |
3130 | /* for a description of osdmap versions, and when they were introduced, please |
3131 | * refer to | |
3132 | * doc/dev/osd_internals/osdmap_versions.txt | |
3133 | */ | |
9f95a23c | 3134 | void OSDMap::decode(ceph::buffer::list& bl) |
7c673cae | 3135 | { |
11fdf7f2 | 3136 | auto p = bl.cbegin(); |
7c673cae FG |
3137 | decode(p); |
3138 | } | |
3139 | ||
9f95a23c | 3140 | void OSDMap::decode_classic(ceph::buffer::list::const_iterator& p) |
7c673cae | 3141 | { |
11fdf7f2 | 3142 | using ceph::decode; |
7c673cae FG |
3143 | __u32 n, t; |
3144 | __u16 v; | |
11fdf7f2 | 3145 | decode(v, p); |
7c673cae FG |
3146 | |
3147 | // base | |
11fdf7f2 TL |
3148 | decode(fsid, p); |
3149 | decode(epoch, p); | |
3150 | decode(created, p); | |
3151 | decode(modified, p); | |
7c673cae FG |
3152 | |
3153 | if (v < 6) { | |
3154 | if (v < 4) { | |
3155 | int32_t max_pools = 0; | |
11fdf7f2 | 3156 | decode(max_pools, p); |
7c673cae FG |
3157 | pool_max = max_pools; |
3158 | } | |
3159 | pools.clear(); | |
11fdf7f2 | 3160 | decode(n, p); |
7c673cae | 3161 | while (n--) { |
11fdf7f2 TL |
3162 | decode(t, p); |
3163 | decode(pools[t], p); | |
7c673cae FG |
3164 | } |
3165 | if (v == 4) { | |
11fdf7f2 | 3166 | decode(n, p); |
7c673cae FG |
3167 | pool_max = n; |
3168 | } else if (v == 5) { | |
3169 | pool_name.clear(); | |
11fdf7f2 | 3170 | decode(n, p); |
7c673cae | 3171 | while (n--) { |
11fdf7f2 TL |
3172 | decode(t, p); |
3173 | decode(pool_name[t], p); | |
7c673cae | 3174 | } |
11fdf7f2 | 3175 | decode(n, p); |
7c673cae FG |
3176 | pool_max = n; |
3177 | } | |
3178 | } else { | |
11fdf7f2 TL |
3179 | decode(pools, p); |
3180 | decode(pool_name, p); | |
3181 | decode(pool_max, p); | |
7c673cae FG |
3182 | } |
3183 | // kludge around some old bug that zeroed out pool_max (#2307) | |
3184 | if (pools.size() && pool_max < pools.rbegin()->first) { | |
3185 | pool_max = pools.rbegin()->first; | |
3186 | } | |
3187 | ||
11fdf7f2 | 3188 | decode(flags, p); |
7c673cae | 3189 | |
11fdf7f2 | 3190 | decode(max_osd, p); |
31f18b77 FG |
3191 | { |
3192 | vector<uint8_t> os; | |
11fdf7f2 | 3193 | decode(os, p); |
31f18b77 FG |
3194 | osd_state.resize(os.size()); |
3195 | for (unsigned i = 0; i < os.size(); ++i) { | |
3196 | osd_state[i] = os[i]; | |
3197 | } | |
3198 | } | |
11fdf7f2 TL |
3199 | decode(osd_weight, p); |
3200 | decode(osd_addrs->client_addrs, p); | |
7c673cae FG |
3201 | if (v <= 5) { |
3202 | pg_temp->clear(); | |
11fdf7f2 | 3203 | decode(n, p); |
7c673cae FG |
3204 | while (n--) { |
3205 | old_pg_t opg; | |
9f95a23c | 3206 | ceph::decode_raw(opg, p); |
31f18b77 | 3207 | mempool::osdmap::vector<int32_t> v; |
11fdf7f2 | 3208 | decode(v, p); |
31f18b77 | 3209 | pg_temp->set(pg_t(opg), v); |
7c673cae FG |
3210 | } |
3211 | } else { | |
11fdf7f2 | 3212 | decode(*pg_temp, p); |
7c673cae FG |
3213 | } |
3214 | ||
3215 | // crush | |
9f95a23c | 3216 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3217 | decode(cbl, p); |
3218 | auto cblp = cbl.cbegin(); | |
7c673cae FG |
3219 | crush->decode(cblp); |
3220 | ||
3221 | // extended | |
3222 | __u16 ev = 0; | |
3223 | if (v >= 5) | |
11fdf7f2 TL |
3224 | decode(ev, p); |
3225 | decode(osd_addrs->hb_back_addrs, p); | |
3226 | decode(osd_info, p); | |
7c673cae | 3227 | if (v < 5) |
11fdf7f2 | 3228 | decode(pool_name, p); |
7c673cae | 3229 | |
f67539c2 | 3230 | decode(blocklist, p); |
7c673cae | 3231 | if (ev >= 6) |
11fdf7f2 | 3232 | decode(osd_addrs->cluster_addrs, p); |
7c673cae | 3233 | else |
11fdf7f2 | 3234 | osd_addrs->cluster_addrs.resize(osd_addrs->client_addrs.size()); |
7c673cae FG |
3235 | |
3236 | if (ev >= 7) { | |
11fdf7f2 TL |
3237 | decode(cluster_snapshot_epoch, p); |
3238 | decode(cluster_snapshot, p); | |
7c673cae FG |
3239 | } |
3240 | ||
3241 | if (ev >= 8) { | |
11fdf7f2 | 3242 | decode(*osd_uuid, p); |
7c673cae FG |
3243 | } else { |
3244 | osd_uuid->resize(max_osd); | |
3245 | } | |
3246 | if (ev >= 9) | |
11fdf7f2 | 3247 | decode(osd_xinfo, p); |
7c673cae FG |
3248 | else |
3249 | osd_xinfo.resize(max_osd); | |
3250 | ||
3251 | if (ev >= 10) | |
11fdf7f2 | 3252 | decode(osd_addrs->hb_front_addrs, p); |
7c673cae | 3253 | else |
11fdf7f2 | 3254 | osd_addrs->hb_front_addrs.resize(osd_addrs->hb_back_addrs.size()); |
7c673cae FG |
3255 | |
3256 | osd_primary_affinity.reset(); | |
3257 | ||
3258 | post_decode(); | |
3259 | } | |
3260 | ||
9f95a23c | 3261 | void OSDMap::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 3262 | { |
11fdf7f2 | 3263 | using ceph::decode; |
7c673cae FG |
3264 | /** |
3265 | * Older encodings of the OSDMap had a single struct_v which | |
3266 | * covered the whole encoding, and was prior to our modern | |
3267 | * stuff which includes a compatv and a size. So if we see | |
3268 | * a struct_v < 7, we must rewind to the beginning and use our | |
3269 | * classic decoder. | |
3270 | */ | |
3271 | size_t start_offset = bl.get_off(); | |
3272 | size_t tail_offset = 0; | |
9f95a23c | 3273 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
3274 | |
3275 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
3276 | if (struct_v < 7) { | |
11fdf7f2 | 3277 | bl.seek(start_offset); |
7c673cae FG |
3278 | decode_classic(bl); |
3279 | return; | |
3280 | } | |
3281 | /** | |
3282 | * Since we made it past that hurdle, we can use our normal paths. | |
3283 | */ | |
3284 | { | |
11fdf7f2 | 3285 | DECODE_START(9, bl); // client-usable data |
7c673cae | 3286 | // base |
11fdf7f2 TL |
3287 | decode(fsid, bl); |
3288 | decode(epoch, bl); | |
3289 | decode(created, bl); | |
3290 | decode(modified, bl); | |
7c673cae | 3291 | |
11fdf7f2 TL |
3292 | decode(pools, bl); |
3293 | decode(pool_name, bl); | |
3294 | decode(pool_max, bl); | |
7c673cae | 3295 | |
11fdf7f2 | 3296 | decode(flags, bl); |
7c673cae | 3297 | |
11fdf7f2 | 3298 | decode(max_osd, bl); |
31f18b77 | 3299 | if (struct_v >= 5) { |
11fdf7f2 | 3300 | decode(osd_state, bl); |
31f18b77 FG |
3301 | } else { |
3302 | vector<uint8_t> os; | |
11fdf7f2 | 3303 | decode(os, bl); |
31f18b77 FG |
3304 | osd_state.resize(os.size()); |
3305 | for (unsigned i = 0; i < os.size(); ++i) { | |
3306 | osd_state[i] = os[i]; | |
3307 | } | |
3308 | } | |
11fdf7f2 TL |
3309 | decode(osd_weight, bl); |
3310 | decode(osd_addrs->client_addrs, bl); | |
7c673cae | 3311 | |
11fdf7f2 TL |
3312 | decode(*pg_temp, bl); |
3313 | decode(*primary_temp, bl); | |
3314 | // dates back to firefly. version increased from 2 to 3 still in firefly. | |
3315 | // do we really still need to keep this around? even for old clients? | |
7c673cae FG |
3316 | if (struct_v >= 2) { |
3317 | osd_primary_affinity.reset(new mempool::osdmap::vector<__u32>); | |
11fdf7f2 | 3318 | decode(*osd_primary_affinity, bl); |
7c673cae FG |
3319 | if (osd_primary_affinity->empty()) |
3320 | osd_primary_affinity.reset(); | |
3321 | } else { | |
3322 | osd_primary_affinity.reset(); | |
3323 | } | |
3324 | ||
3325 | // crush | |
9f95a23c | 3326 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3327 | decode(cbl, bl); |
3328 | auto cblp = cbl.cbegin(); | |
7c673cae | 3329 | crush->decode(cblp); |
11fdf7f2 TL |
3330 | // added in firefly; version increased in luminous, so it affects |
3331 | // giant, hammer, infernallis, jewel, and kraken. probably should be left | |
3332 | // alone until we require clients to be all luminous? | |
7c673cae | 3333 | if (struct_v >= 3) { |
11fdf7f2 | 3334 | decode(erasure_code_profiles, bl); |
7c673cae FG |
3335 | } else { |
3336 | erasure_code_profiles.clear(); | |
3337 | } | |
11fdf7f2 TL |
3338 | // version increased from 3 to 4 still in luminous, so same as above |
3339 | // applies. | |
7c673cae | 3340 | if (struct_v >= 4) { |
11fdf7f2 TL |
3341 | decode(pg_upmap, bl); |
3342 | decode(pg_upmap_items, bl); | |
7c673cae FG |
3343 | } else { |
3344 | pg_upmap.clear(); | |
3345 | pg_upmap_items.clear(); | |
3346 | } | |
11fdf7f2 TL |
3347 | // again, version increased from 5 to 6 still in luminous, so above |
3348 | // applies. | |
31f18b77 | 3349 | if (struct_v >= 6) { |
11fdf7f2 TL |
3350 | decode(crush_version, bl); |
3351 | } | |
3352 | // version increase from 6 to 7 in mimic | |
3353 | if (struct_v >= 7) { | |
3354 | decode(new_removed_snaps, bl); | |
3355 | decode(new_purged_snaps, bl); | |
3356 | } | |
3357 | // version increase from 7 to 8, 8 to 9, in nautilus. | |
3358 | if (struct_v >= 9) { | |
3359 | decode(last_up_change, bl); | |
3360 | decode(last_in_change, bl); | |
31f18b77 | 3361 | } |
7c673cae FG |
3362 | DECODE_FINISH(bl); // client-usable data |
3363 | } | |
3364 | ||
3365 | { | |
f67539c2 | 3366 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
3367 | decode(osd_addrs->hb_back_addrs, bl); |
3368 | decode(osd_info, bl); | |
f67539c2 | 3369 | decode(blocklist, bl); |
11fdf7f2 TL |
3370 | decode(osd_addrs->cluster_addrs, bl); |
3371 | decode(cluster_snapshot_epoch, bl); | |
3372 | decode(cluster_snapshot, bl); | |
3373 | decode(*osd_uuid, bl); | |
3374 | decode(osd_xinfo, bl); | |
3375 | decode(osd_addrs->hb_front_addrs, bl); | |
3376 | // | |
7c673cae | 3377 | if (struct_v >= 2) { |
11fdf7f2 TL |
3378 | decode(nearfull_ratio, bl); |
3379 | decode(full_ratio, bl); | |
7c673cae FG |
3380 | } else { |
3381 | nearfull_ratio = 0; | |
3382 | full_ratio = 0; | |
3383 | } | |
3384 | if (struct_v >= 3) { | |
11fdf7f2 | 3385 | decode(backfillfull_ratio, bl); |
7c673cae FG |
3386 | } else { |
3387 | backfillfull_ratio = 0; | |
3388 | } | |
31f18b77 FG |
3389 | if (struct_v == 4) { |
3390 | string r; | |
11fdf7f2 | 3391 | decode(r, bl); |
31f18b77 FG |
3392 | if (r.length()) |
3393 | require_min_compat_client = ceph_release_from_name(r.c_str()); | |
3394 | } | |
3395 | if (struct_v >= 5) { | |
11fdf7f2 TL |
3396 | decode(require_min_compat_client, bl); |
3397 | decode(require_osd_release, bl); | |
9f95a23c | 3398 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
3399 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
3400 | } | |
9f95a23c | 3401 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 3402 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3403 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
3404 | } |
3405 | } else { | |
3406 | if (flags & CEPH_OSDMAP_REQUIRE_LUMINOUS) { | |
3407 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 3408 | require_osd_release = ceph_release_t::luminous; |
31f18b77 | 3409 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3410 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 | 3411 | } else if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { |
9f95a23c | 3412 | require_osd_release = ceph_release_t::kraken; |
31f18b77 | 3413 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { |
9f95a23c | 3414 | require_osd_release = ceph_release_t::jewel; |
31f18b77 | 3415 | } else { |
9f95a23c | 3416 | require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
3417 | } |
3418 | } | |
11fdf7f2 TL |
3419 | if (struct_v >= 6) { |
3420 | decode(removed_snaps_queue, bl); | |
3421 | } | |
81eedcae TL |
3422 | if (struct_v >= 8) { |
3423 | decode(crush_node_flags, bl); | |
3424 | } else { | |
3425 | crush_node_flags.clear(); | |
3426 | } | |
3427 | if (struct_v >= 9) { | |
3428 | decode(device_class_flags, bl); | |
3429 | } else { | |
3430 | device_class_flags.clear(); | |
3431 | } | |
f67539c2 TL |
3432 | if (struct_v >= 10) { |
3433 | decode(stretch_mode_enabled, bl); | |
3434 | decode(stretch_bucket_count, bl); | |
3435 | decode(degraded_stretch_mode, bl); | |
3436 | decode(recovering_stretch_mode, bl); | |
3437 | decode(stretch_mode_bucket, bl); | |
3438 | } else { | |
3439 | stretch_mode_enabled = false; | |
3440 | stretch_bucket_count = 0; | |
3441 | degraded_stretch_mode = 0; | |
3442 | recovering_stretch_mode = 0; | |
3443 | stretch_mode_bucket = 0; | |
3444 | } | |
7c673cae FG |
3445 | DECODE_FINISH(bl); // osd-only data |
3446 | } | |
3447 | ||
3448 | if (struct_v >= 8) { | |
3449 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 3450 | decode(crc, bl); |
7c673cae FG |
3451 | tail_offset = bl.get_off(); |
3452 | crc_defined = true; | |
3453 | } else { | |
3454 | crc_defined = false; | |
3455 | crc = 0; | |
3456 | } | |
3457 | ||
3458 | DECODE_FINISH(bl); // wrapper | |
3459 | ||
3460 | if (tail_offset) { | |
3461 | // verify crc | |
3462 | uint32_t actual = crc_front.crc32c(-1); | |
3463 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 3464 | ceph::buffer::list tail; |
7c673cae FG |
3465 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
3466 | actual = tail.crc32c(actual); | |
3467 | } | |
3468 | if (crc != actual) { | |
3469 | ostringstream ss; | |
3470 | ss << "bad crc, actual " << actual << " != expected " << crc; | |
3471 | string s = ss.str(); | |
9f95a23c | 3472 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
3473 | } |
3474 | } | |
3475 | ||
3476 | post_decode(); | |
3477 | } | |
3478 | ||
3479 | void OSDMap::post_decode() | |
3480 | { | |
3481 | // index pool names | |
3482 | name_pool.clear(); | |
3483 | for (const auto &pname : pool_name) { | |
3484 | name_pool[pname.second] = pname.first; | |
3485 | } | |
3486 | ||
3487 | calc_num_osds(); | |
3488 | _calc_up_osd_features(); | |
3489 | } | |
3490 | ||
3491 | void OSDMap::dump_erasure_code_profiles( | |
3492 | const mempool::osdmap::map<string,map<string,string>>& profiles, | |
3493 | Formatter *f) | |
3494 | { | |
3495 | f->open_object_section("erasure_code_profiles"); | |
3496 | for (const auto &profile : profiles) { | |
3497 | f->open_object_section(profile.first.c_str()); | |
3498 | for (const auto &profm : profile.second) { | |
9f95a23c | 3499 | f->dump_string(profm.first.c_str(), profm.second); |
7c673cae FG |
3500 | } |
3501 | f->close_section(); | |
3502 | } | |
3503 | f->close_section(); | |
3504 | } | |
3505 | ||
9f95a23c TL |
3506 | void OSDMap::dump_osds(Formatter *f) const |
3507 | { | |
3508 | f->open_array_section("osds"); | |
3509 | for (int i=0; i<get_max_osd(); i++) { | |
3510 | if (exists(i)) { | |
3511 | dump_osd(i, f); | |
3512 | } | |
3513 | } | |
3514 | f->close_section(); | |
3515 | } | |
3516 | ||
3517 | void OSDMap::dump_osd(int id, Formatter *f) const | |
3518 | { | |
3519 | ceph_assert(f != nullptr); | |
3520 | if (!exists(id)) { | |
3521 | return; | |
3522 | } | |
3523 | ||
3524 | f->open_object_section("osd_info"); | |
3525 | f->dump_int("osd", id); | |
3526 | f->dump_stream("uuid") << get_uuid(id); | |
3527 | f->dump_int("up", is_up(id)); | |
3528 | f->dump_int("in", is_in(id)); | |
3529 | f->dump_float("weight", get_weightf(id)); | |
3530 | f->dump_float("primary_affinity", get_primary_affinityf(id)); | |
3531 | get_info(id).dump(f); | |
3532 | f->dump_object("public_addrs", get_addrs(id)); | |
3533 | f->dump_object("cluster_addrs", get_cluster_addrs(id)); | |
3534 | f->dump_object("heartbeat_back_addrs", get_hb_back_addrs(id)); | |
3535 | f->dump_object("heartbeat_front_addrs", get_hb_front_addrs(id)); | |
3536 | // compat | |
3537 | f->dump_stream("public_addr") << get_addrs(id).get_legacy_str(); | |
3538 | f->dump_stream("cluster_addr") << get_cluster_addrs(id).get_legacy_str(); | |
3539 | f->dump_stream("heartbeat_back_addr") | |
3540 | << get_hb_back_addrs(id).get_legacy_str(); | |
3541 | f->dump_stream("heartbeat_front_addr") | |
3542 | << get_hb_front_addrs(id).get_legacy_str(); | |
3543 | ||
3544 | set<string> st; | |
3545 | get_state(id, st); | |
3546 | f->open_array_section("state"); | |
3547 | for (const auto &state : st) | |
3548 | f->dump_string("state", state); | |
3549 | f->close_section(); | |
3550 | ||
3551 | f->close_section(); | |
3552 | } | |
3553 | ||
7c673cae FG |
3554 | void OSDMap::dump(Formatter *f) const |
3555 | { | |
3556 | f->dump_int("epoch", get_epoch()); | |
3557 | f->dump_stream("fsid") << get_fsid(); | |
3558 | f->dump_stream("created") << get_created(); | |
3559 | f->dump_stream("modified") << get_modified(); | |
11fdf7f2 TL |
3560 | f->dump_stream("last_up_change") << last_up_change; |
3561 | f->dump_stream("last_in_change") << last_in_change; | |
7c673cae | 3562 | f->dump_string("flags", get_flag_string()); |
11fdf7f2 TL |
3563 | f->dump_unsigned("flags_num", flags); |
3564 | f->open_array_section("flags_set"); | |
3565 | set<string> flagset; | |
3566 | get_flag_set(&flagset); | |
3567 | for (auto p : flagset) { | |
3568 | f->dump_string("flag", p); | |
3569 | } | |
3570 | f->close_section(); | |
31f18b77 | 3571 | f->dump_unsigned("crush_version", get_crush_version()); |
7c673cae FG |
3572 | f->dump_float("full_ratio", full_ratio); |
3573 | f->dump_float("backfillfull_ratio", backfillfull_ratio); | |
3574 | f->dump_float("nearfull_ratio", nearfull_ratio); | |
3575 | f->dump_string("cluster_snapshot", get_cluster_snapshot()); | |
3576 | f->dump_int("pool_max", get_pool_max()); | |
3577 | f->dump_int("max_osd", get_max_osd()); | |
31f18b77 | 3578 | f->dump_string("require_min_compat_client", |
f67539c2 | 3579 | to_string(require_min_compat_client)); |
31f18b77 | 3580 | f->dump_string("min_compat_client", |
f67539c2 | 3581 | to_string(get_min_compat_client())); |
31f18b77 | 3582 | f->dump_string("require_osd_release", |
f67539c2 | 3583 | to_string(require_osd_release)); |
7c673cae FG |
3584 | |
3585 | f->open_array_section("pools"); | |
3586 | for (const auto &pool : pools) { | |
3587 | std::string name("<unknown>"); | |
3588 | const auto &pni = pool_name.find(pool.first); | |
3589 | if (pni != pool_name.end()) | |
3590 | name = pni->second; | |
3591 | f->open_object_section("pool"); | |
3592 | f->dump_int("pool", pool.first); | |
3593 | f->dump_string("pool_name", name); | |
3594 | pool.second.dump(f); | |
3595 | f->close_section(); | |
3596 | } | |
3597 | f->close_section(); | |
3598 | ||
9f95a23c | 3599 | dump_osds(f); |
7c673cae FG |
3600 | |
3601 | f->open_array_section("osd_xinfo"); | |
3602 | for (int i=0; i<get_max_osd(); i++) { | |
3603 | if (exists(i)) { | |
3604 | f->open_object_section("xinfo"); | |
3605 | f->dump_int("osd", i); | |
3606 | osd_xinfo[i].dump(f); | |
3607 | f->close_section(); | |
3608 | } | |
3609 | } | |
3610 | f->close_section(); | |
3611 | ||
3612 | f->open_array_section("pg_upmap"); | |
3613 | for (auto& p : pg_upmap) { | |
3614 | f->open_object_section("mapping"); | |
3615 | f->dump_stream("pgid") << p.first; | |
3616 | f->open_array_section("osds"); | |
3617 | for (auto q : p.second) { | |
3618 | f->dump_int("osd", q); | |
3619 | } | |
3620 | f->close_section(); | |
3621 | f->close_section(); | |
3622 | } | |
3623 | f->close_section(); | |
3624 | f->open_array_section("pg_upmap_items"); | |
3625 | for (auto& p : pg_upmap_items) { | |
3626 | f->open_object_section("mapping"); | |
3627 | f->dump_stream("pgid") << p.first; | |
3628 | f->open_array_section("mappings"); | |
3629 | for (auto& q : p.second) { | |
3630 | f->open_object_section("mapping"); | |
3631 | f->dump_int("from", q.first); | |
3632 | f->dump_int("to", q.second); | |
3633 | f->close_section(); | |
3634 | } | |
3635 | f->close_section(); | |
3636 | f->close_section(); | |
3637 | } | |
3638 | f->close_section(); | |
3639 | f->open_array_section("pg_temp"); | |
31f18b77 | 3640 | pg_temp->dump(f); |
7c673cae FG |
3641 | f->close_section(); |
3642 | ||
3643 | f->open_array_section("primary_temp"); | |
3644 | for (const auto &pg : *primary_temp) { | |
3645 | f->dump_stream("pgid") << pg.first; | |
3646 | f->dump_int("osd", pg.second); | |
3647 | } | |
3648 | f->close_section(); // primary_temp | |
3649 | ||
f67539c2 TL |
3650 | f->open_object_section("blocklist"); |
3651 | for (const auto &addr : blocklist) { | |
7c673cae FG |
3652 | stringstream ss; |
3653 | ss << addr.first; | |
3654 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3655 | } | |
3656 | f->close_section(); | |
3657 | ||
3658 | dump_erasure_code_profiles(erasure_code_profiles, f); | |
11fdf7f2 TL |
3659 | |
3660 | f->open_array_section("removed_snaps_queue"); | |
3661 | for (auto& p : removed_snaps_queue) { | |
3662 | f->open_object_section("pool"); | |
3663 | f->dump_int("pool", p.first); | |
3664 | f->open_array_section("snaps"); | |
3665 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3666 | f->open_object_section("interval"); | |
3667 | f->dump_unsigned("begin", q.get_start()); | |
3668 | f->dump_unsigned("length", q.get_len()); | |
3669 | f->close_section(); | |
3670 | } | |
3671 | f->close_section(); | |
3672 | f->close_section(); | |
3673 | } | |
3674 | f->close_section(); | |
3675 | f->open_array_section("new_removed_snaps"); | |
3676 | for (auto& p : new_removed_snaps) { | |
3677 | f->open_object_section("pool"); | |
3678 | f->dump_int("pool", p.first); | |
3679 | f->open_array_section("snaps"); | |
3680 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3681 | f->open_object_section("interval"); | |
3682 | f->dump_unsigned("begin", q.get_start()); | |
3683 | f->dump_unsigned("length", q.get_len()); | |
3684 | f->close_section(); | |
3685 | } | |
3686 | f->close_section(); | |
3687 | f->close_section(); | |
3688 | } | |
3689 | f->close_section(); | |
3690 | f->open_array_section("new_purged_snaps"); | |
3691 | for (auto& p : new_purged_snaps) { | |
3692 | f->open_object_section("pool"); | |
3693 | f->dump_int("pool", p.first); | |
3694 | f->open_array_section("snaps"); | |
3695 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3696 | f->open_object_section("interval"); | |
3697 | f->dump_unsigned("begin", q.get_start()); | |
3698 | f->dump_unsigned("length", q.get_len()); | |
3699 | f->close_section(); | |
3700 | } | |
3701 | f->close_section(); | |
3702 | f->close_section(); | |
3703 | } | |
3704 | f->close_section(); | |
81eedcae TL |
3705 | f->open_object_section("crush_node_flags"); |
3706 | for (auto& i : crush_node_flags) { | |
3707 | string s = crush->item_exists(i.first) ? crush->get_item_name(i.first) | |
3708 | : stringify(i.first); | |
3709 | f->open_array_section(s.c_str()); | |
3710 | set<string> st; | |
3711 | calc_state_set(i.second, st); | |
3712 | for (auto& j : st) { | |
3713 | f->dump_string("flag", j); | |
3714 | } | |
3715 | f->close_section(); | |
3716 | } | |
3717 | f->close_section(); | |
3718 | f->open_object_section("device_class_flags"); | |
3719 | for (auto& i : device_class_flags) { | |
3720 | const char* class_name = crush->get_class_name(i.first); | |
3721 | string s = class_name ? class_name : stringify(i.first); | |
3722 | f->open_array_section(s.c_str()); | |
3723 | set<string> st; | |
3724 | calc_state_set(i.second, st); | |
3725 | for (auto& j : st) { | |
3726 | f->dump_string("flag", j); | |
3727 | } | |
3728 | f->close_section(); | |
3729 | } | |
3730 | f->close_section(); | |
f67539c2 TL |
3731 | f->open_object_section("stretch_mode"); |
3732 | { | |
3733 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
3734 | f->dump_unsigned("stretch_bucket_count", stretch_bucket_count); | |
3735 | f->dump_unsigned("degraded_stretch_mode", degraded_stretch_mode); | |
3736 | f->dump_unsigned("recovering_stretch_mode", recovering_stretch_mode); | |
3737 | f->dump_int("stretch_mode_bucket", stretch_mode_bucket); | |
3738 | } | |
3739 | f->close_section(); | |
7c673cae FG |
3740 | } |
3741 | ||
3742 | void OSDMap::generate_test_instances(list<OSDMap*>& o) | |
3743 | { | |
3744 | o.push_back(new OSDMap); | |
3745 | ||
3746 | CephContext *cct = new CephContext(CODE_ENVIRONMENT_UTILITY); | |
3747 | o.push_back(new OSDMap); | |
3748 | uuid_d fsid; | |
224ce89b | 3749 | o.back()->build_simple(cct, 1, fsid, 16); |
7c673cae | 3750 | o.back()->created = o.back()->modified = utime_t(1, 2); // fix timestamp |
f67539c2 | 3751 | o.back()->blocklist[entity_addr_t()] = utime_t(5, 6); |
7c673cae FG |
3752 | cct->put(); |
3753 | } | |
3754 | ||
3755 | string OSDMap::get_flag_string(unsigned f) | |
3756 | { | |
3757 | string s; | |
7c673cae FG |
3758 | if (f & CEPH_OSDMAP_PAUSERD) |
3759 | s += ",pauserd"; | |
3760 | if (f & CEPH_OSDMAP_PAUSEWR) | |
3761 | s += ",pausewr"; | |
3762 | if (f & CEPH_OSDMAP_PAUSEREC) | |
3763 | s += ",pauserec"; | |
3764 | if (f & CEPH_OSDMAP_NOUP) | |
3765 | s += ",noup"; | |
3766 | if (f & CEPH_OSDMAP_NODOWN) | |
3767 | s += ",nodown"; | |
3768 | if (f & CEPH_OSDMAP_NOOUT) | |
3769 | s += ",noout"; | |
3770 | if (f & CEPH_OSDMAP_NOIN) | |
3771 | s += ",noin"; | |
3772 | if (f & CEPH_OSDMAP_NOBACKFILL) | |
3773 | s += ",nobackfill"; | |
3774 | if (f & CEPH_OSDMAP_NOREBALANCE) | |
3775 | s += ",norebalance"; | |
3776 | if (f & CEPH_OSDMAP_NORECOVER) | |
3777 | s += ",norecover"; | |
3778 | if (f & CEPH_OSDMAP_NOSCRUB) | |
3779 | s += ",noscrub"; | |
3780 | if (f & CEPH_OSDMAP_NODEEP_SCRUB) | |
3781 | s += ",nodeep-scrub"; | |
3782 | if (f & CEPH_OSDMAP_NOTIERAGENT) | |
3783 | s += ",notieragent"; | |
11fdf7f2 TL |
3784 | if (f & CEPH_OSDMAP_NOSNAPTRIM) |
3785 | s += ",nosnaptrim"; | |
7c673cae FG |
3786 | if (f & CEPH_OSDMAP_SORTBITWISE) |
3787 | s += ",sortbitwise"; | |
3788 | if (f & CEPH_OSDMAP_REQUIRE_JEWEL) | |
3789 | s += ",require_jewel_osds"; | |
3790 | if (f & CEPH_OSDMAP_REQUIRE_KRAKEN) | |
3791 | s += ",require_kraken_osds"; | |
3792 | if (f & CEPH_OSDMAP_REQUIRE_LUMINOUS) | |
3793 | s += ",require_luminous_osds"; | |
c07f9fc5 FG |
3794 | if (f & CEPH_OSDMAP_RECOVERY_DELETES) |
3795 | s += ",recovery_deletes"; | |
181888fb FG |
3796 | if (f & CEPH_OSDMAP_PURGED_SNAPDIRS) |
3797 | s += ",purged_snapdirs"; | |
f64942e4 AA |
3798 | if (f & CEPH_OSDMAP_PGLOG_HARDLIMIT) |
3799 | s += ",pglog_hardlimit"; | |
7c673cae FG |
3800 | if (s.length()) |
3801 | s.erase(0, 1); | |
3802 | return s; | |
3803 | } | |
3804 | ||
3805 | string OSDMap::get_flag_string() const | |
3806 | { | |
3807 | return get_flag_string(flags); | |
3808 | } | |
3809 | ||
7c673cae FG |
3810 | void OSDMap::print_pools(ostream& out) const |
3811 | { | |
3812 | for (const auto &pool : pools) { | |
3813 | std::string name("<unknown>"); | |
3814 | const auto &pni = pool_name.find(pool.first); | |
3815 | if (pni != pool_name.end()) | |
3816 | name = pni->second; | |
3817 | out << "pool " << pool.first | |
3818 | << " '" << name | |
3819 | << "' " << pool.second << "\n"; | |
3820 | ||
3821 | for (const auto &snap : pool.second.snaps) | |
3822 | out << "\tsnap " << snap.second.snapid << " '" << snap.second.name << "' " << snap.second.stamp << "\n"; | |
3823 | ||
3824 | if (!pool.second.removed_snaps.empty()) | |
3825 | out << "\tremoved_snaps " << pool.second.removed_snaps << "\n"; | |
11fdf7f2 TL |
3826 | auto p = removed_snaps_queue.find(pool.first); |
3827 | if (p != removed_snaps_queue.end()) { | |
3828 | out << "\tremoved_snaps_queue " << p->second << "\n"; | |
3829 | } | |
7c673cae FG |
3830 | } |
3831 | out << std::endl; | |
3832 | } | |
3833 | ||
9f95a23c TL |
3834 | void OSDMap::print_osds(ostream& out) const |
3835 | { | |
3836 | for (int i=0; i<get_max_osd(); i++) { | |
3837 | if (exists(i)) { | |
3838 | print_osd(i, out); | |
3839 | } | |
3840 | } | |
3841 | } | |
3842 | void OSDMap::print_osd(int id, ostream& out) const | |
3843 | { | |
3844 | if (!exists(id)) { | |
3845 | return; | |
3846 | } | |
3847 | ||
3848 | out << "osd." << id; | |
3849 | out << (is_up(id) ? " up ":" down"); | |
3850 | out << (is_in(id) ? " in ":" out"); | |
3851 | out << " weight " << get_weightf(id); | |
3852 | if (get_primary_affinity(id) != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
3853 | out << " primary_affinity " << get_primary_affinityf(id); | |
3854 | } | |
3855 | const osd_info_t& info(get_info(id)); | |
3856 | out << " " << info; | |
3857 | out << " " << get_addrs(id) << " " << get_cluster_addrs(id); | |
3858 | set<string> st; | |
3859 | get_state(id, st); | |
3860 | out << " " << st; | |
3861 | if (!get_uuid(id).is_zero()) { | |
3862 | out << " " << get_uuid(id); | |
3863 | } | |
3864 | out << "\n"; | |
3865 | } | |
3866 | ||
7c673cae FG |
3867 | void OSDMap::print(ostream& out) const |
3868 | { | |
3869 | out << "epoch " << get_epoch() << "\n" | |
3870 | << "fsid " << get_fsid() << "\n" | |
3871 | << "created " << get_created() << "\n" | |
3872 | << "modified " << get_modified() << "\n"; | |
3873 | ||
3874 | out << "flags " << get_flag_string() << "\n"; | |
31f18b77 | 3875 | out << "crush_version " << get_crush_version() << "\n"; |
7c673cae FG |
3876 | out << "full_ratio " << full_ratio << "\n"; |
3877 | out << "backfillfull_ratio " << backfillfull_ratio << "\n"; | |
3878 | out << "nearfull_ratio " << nearfull_ratio << "\n"; | |
9f95a23c | 3879 | if (require_min_compat_client != ceph_release_t::unknown) { |
31f18b77 | 3880 | out << "require_min_compat_client " |
9f95a23c | 3881 | << require_min_compat_client << "\n"; |
7c673cae | 3882 | } |
9f95a23c | 3883 | out << "min_compat_client " << get_min_compat_client() |
31f18b77 | 3884 | << "\n"; |
9f95a23c TL |
3885 | if (require_osd_release > ceph_release_t::unknown) { |
3886 | out << "require_osd_release " << require_osd_release | |
224ce89b WB |
3887 | << "\n"; |
3888 | } | |
f67539c2 TL |
3889 | out << "stretch_mode_enabled " << (stretch_mode_enabled ? "true" : "false") << "\n"; |
3890 | if (stretch_mode_enabled) { | |
3891 | out << "stretch_bucket_count " << stretch_bucket_count << "\n"; | |
3892 | out << "degraded_stretch_mode " << degraded_stretch_mode << "\n"; | |
3893 | out << "recovering_stretch_mode " << recovering_stretch_mode << "\n"; | |
3894 | out << "stretch_mode_bucket " << stretch_mode_bucket << "\n"; | |
3895 | } | |
7c673cae FG |
3896 | if (get_cluster_snapshot().length()) |
3897 | out << "cluster_snapshot " << get_cluster_snapshot() << "\n"; | |
3898 | out << "\n"; | |
3899 | ||
3900 | print_pools(out); | |
3901 | ||
3902 | out << "max_osd " << get_max_osd() << "\n"; | |
9f95a23c | 3903 | print_osds(out); |
7c673cae FG |
3904 | out << std::endl; |
3905 | ||
3906 | for (auto& p : pg_upmap) { | |
3907 | out << "pg_upmap " << p.first << " " << p.second << "\n"; | |
3908 | } | |
3909 | for (auto& p : pg_upmap_items) { | |
3910 | out << "pg_upmap_items " << p.first << " " << p.second << "\n"; | |
3911 | } | |
3912 | ||
f67539c2 | 3913 | for (const auto& pg : *pg_temp) |
7c673cae FG |
3914 | out << "pg_temp " << pg.first << " " << pg.second << "\n"; |
3915 | ||
f67539c2 | 3916 | for (const auto& pg : *primary_temp) |
7c673cae FG |
3917 | out << "primary_temp " << pg.first << " " << pg.second << "\n"; |
3918 | ||
f67539c2 TL |
3919 | for (const auto &addr : blocklist) |
3920 | out << "blocklist " << addr.first << " expires " << addr.second << "\n"; | |
7c673cae FG |
3921 | } |
3922 | ||
3923 | class OSDTreePlainDumper : public CrushTreeDumper::Dumper<TextTable> { | |
3924 | public: | |
3925 | typedef CrushTreeDumper::Dumper<TextTable> Parent; | |
31f18b77 FG |
3926 | |
3927 | OSDTreePlainDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
3928 | unsigned f) | |
c07f9fc5 | 3929 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
3930 | |
3931 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
3932 | if (!filter) { |
3933 | return true; // normal case | |
3934 | } | |
3935 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
3936 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
3937 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
3938 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
3939 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
3940 | return true; | |
31f18b77 | 3941 | } |
c07f9fc5 | 3942 | return false; |
31f18b77 FG |
3943 | } |
3944 | ||
3945 | bool should_dump_empty_bucket() const override { | |
3946 | return !filter; | |
3947 | } | |
7c673cae | 3948 | |
11fdf7f2 | 3949 | void init_table(TextTable *tbl) { |
7c673cae | 3950 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 3951 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
3952 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
3953 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
c07f9fc5 | 3954 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae | 3955 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 3956 | tbl->define_column("PRI-AFF", TextTable::LEFT, TextTable::RIGHT); |
11fdf7f2 TL |
3957 | } |
3958 | void dump(TextTable *tbl, string& bucket) { | |
3959 | init_table(tbl); | |
7c673cae | 3960 | |
11fdf7f2 TL |
3961 | if (!bucket.empty()) { |
3962 | set_root(bucket); | |
3963 | Parent::dump(tbl); | |
3964 | } else { | |
3965 | Parent::dump(tbl); | |
3966 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
3967 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) { | |
3968 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), tbl); | |
3969 | } | |
31f18b77 | 3970 | } |
7c673cae FG |
3971 | } |
3972 | } | |
3973 | ||
3974 | protected: | |
3975 | void dump_item(const CrushTreeDumper::Item &qi, TextTable *tbl) override { | |
224ce89b WB |
3976 | const char *c = crush->get_item_class(qi.id); |
3977 | if (!c) | |
3978 | c = ""; | |
7c673cae | 3979 | *tbl << qi.id |
224ce89b | 3980 | << c |
7c673cae FG |
3981 | << weightf_t(qi.weight); |
3982 | ||
3983 | ostringstream name; | |
3984 | for (int k = 0; k < qi.depth; k++) | |
3985 | name << " "; | |
3986 | if (qi.is_bucket()) { | |
3987 | name << crush->get_type_name(crush->get_bucket_type(qi.id)) << " " | |
3988 | << crush->get_item_name(qi.id); | |
3989 | } else { | |
3990 | name << "osd." << qi.id; | |
3991 | } | |
3992 | *tbl << name.str(); | |
3993 | ||
3994 | if (!qi.is_bucket()) { | |
3995 | if (!osdmap->exists(qi.id)) { | |
3996 | *tbl << "DNE" | |
3997 | << 0; | |
3998 | } else { | |
c07f9fc5 FG |
3999 | string s; |
4000 | if (osdmap->is_up(qi.id)) { | |
4001 | s = "up"; | |
4002 | } else if (osdmap->is_destroyed(qi.id)) { | |
4003 | s = "destroyed"; | |
4004 | } else { | |
4005 | s = "down"; | |
4006 | } | |
4007 | *tbl << s | |
7c673cae FG |
4008 | << weightf_t(osdmap->get_weightf(qi.id)) |
4009 | << weightf_t(osdmap->get_primary_affinityf(qi.id)); | |
4010 | } | |
4011 | } | |
4012 | *tbl << TextTable::endrow; | |
4013 | } | |
4014 | ||
4015 | private: | |
4016 | const OSDMap *osdmap; | |
31f18b77 | 4017 | const unsigned filter; |
7c673cae FG |
4018 | }; |
4019 | ||
4020 | class OSDTreeFormattingDumper : public CrushTreeDumper::FormattingDumper { | |
4021 | public: | |
4022 | typedef CrushTreeDumper::FormattingDumper Parent; | |
4023 | ||
31f18b77 FG |
4024 | OSDTreeFormattingDumper(const CrushWrapper *crush, const OSDMap *osdmap_, |
4025 | unsigned f) | |
c07f9fc5 | 4026 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
4027 | |
4028 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
4029 | if (!filter) { |
4030 | return true; // normal case | |
4031 | } | |
4032 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
4033 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
4034 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
4035 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
4036 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
4037 | return true; | |
31f18b77 | 4038 | } |
c07f9fc5 | 4039 | return false; |
31f18b77 FG |
4040 | } |
4041 | ||
4042 | bool should_dump_empty_bucket() const override { | |
4043 | return !filter; | |
4044 | } | |
7c673cae | 4045 | |
11fdf7f2 TL |
4046 | void dump(Formatter *f, string& bucket) { |
4047 | if (!bucket.empty()) { | |
4048 | set_root(bucket); | |
4049 | f->open_array_section("nodes"); | |
4050 | Parent::dump(f); | |
4051 | f->close_section(); | |
4052 | } else { | |
4053 | f->open_array_section("nodes"); | |
4054 | Parent::dump(f); | |
4055 | f->close_section(); | |
4056 | f->open_array_section("stray"); | |
4057 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4058 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) | |
4059 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); | |
4060 | } | |
4061 | f->close_section(); | |
7c673cae | 4062 | } |
7c673cae FG |
4063 | } |
4064 | ||
4065 | protected: | |
4066 | void dump_item_fields(const CrushTreeDumper::Item &qi, Formatter *f) override { | |
4067 | Parent::dump_item_fields(qi, f); | |
4068 | if (!qi.is_bucket()) | |
4069 | { | |
c07f9fc5 FG |
4070 | string s; |
4071 | if (osdmap->is_up(qi.id)) { | |
4072 | s = "up"; | |
4073 | } else if (osdmap->is_destroyed(qi.id)) { | |
4074 | s = "destroyed"; | |
4075 | } else { | |
4076 | s = "down"; | |
4077 | } | |
7c673cae | 4078 | f->dump_unsigned("exists", (int)osdmap->exists(qi.id)); |
c07f9fc5 | 4079 | f->dump_string("status", s); |
7c673cae FG |
4080 | f->dump_float("reweight", osdmap->get_weightf(qi.id)); |
4081 | f->dump_float("primary_affinity", osdmap->get_primary_affinityf(qi.id)); | |
4082 | } | |
4083 | } | |
4084 | ||
4085 | private: | |
4086 | const OSDMap *osdmap; | |
31f18b77 | 4087 | const unsigned filter; |
7c673cae FG |
4088 | }; |
4089 | ||
11fdf7f2 | 4090 | void OSDMap::print_tree(Formatter *f, ostream *out, unsigned filter, string bucket) const |
7c673cae | 4091 | { |
31f18b77 | 4092 | if (f) { |
11fdf7f2 | 4093 | OSDTreeFormattingDumper(crush.get(), this, filter).dump(f, bucket); |
31f18b77 | 4094 | } else { |
11fdf7f2 | 4095 | ceph_assert(out); |
7c673cae | 4096 | TextTable tbl; |
11fdf7f2 | 4097 | OSDTreePlainDumper(crush.get(), this, filter).dump(&tbl, bucket); |
7c673cae FG |
4098 | *out << tbl; |
4099 | } | |
4100 | } | |
4101 | ||
224ce89b | 4102 | void OSDMap::print_summary(Formatter *f, ostream& out, |
11fdf7f2 | 4103 | const string& prefix, bool extra) const |
7c673cae FG |
4104 | { |
4105 | if (f) { | |
7c673cae FG |
4106 | f->dump_int("epoch", get_epoch()); |
4107 | f->dump_int("num_osds", get_num_osds()); | |
4108 | f->dump_int("num_up_osds", get_num_up_osds()); | |
9f95a23c | 4109 | f->dump_int("osd_up_since", last_up_change.to_msec() / 1000); |
7c673cae | 4110 | f->dump_int("num_in_osds", get_num_in_osds()); |
9f95a23c | 4111 | f->dump_int("osd_in_since", last_in_change.to_msec() / 1000); |
7c673cae | 4112 | f->dump_unsigned("num_remapped_pgs", get_num_pg_temp()); |
7c673cae | 4113 | } else { |
11fdf7f2 | 4114 | utime_t now = ceph_clock_now(); |
31f18b77 | 4115 | out << get_num_osds() << " osds: " |
11fdf7f2 TL |
4116 | << get_num_up_osds() << " up"; |
4117 | if (last_up_change != utime_t()) { | |
4118 | out << " (since " << utimespan_str(now - last_up_change) << ")"; | |
4119 | } | |
4120 | out << ", " << get_num_in_osds() << " in"; | |
4121 | if (last_in_change != utime_t()) { | |
4122 | out << " (since " << utimespan_str(now - last_in_change) << ")"; | |
4123 | } | |
4124 | if (extra) | |
4125 | out << "; epoch: e" << get_epoch(); | |
7c673cae FG |
4126 | if (get_num_pg_temp()) |
4127 | out << "; " << get_num_pg_temp() << " remapped pgs"; | |
4128 | out << "\n"; | |
4129 | uint64_t important_flags = flags & ~CEPH_OSDMAP_SEMIHIDDEN_FLAGS; | |
4130 | if (important_flags) | |
224ce89b | 4131 | out << prefix << "flags " << get_flag_string(important_flags) << "\n"; |
7c673cae FG |
4132 | } |
4133 | } | |
4134 | ||
4135 | void OSDMap::print_oneline_summary(ostream& out) const | |
4136 | { | |
4137 | out << "e" << get_epoch() << ": " | |
31f18b77 | 4138 | << get_num_osds() << " total, " |
7c673cae FG |
4139 | << get_num_up_osds() << " up, " |
4140 | << get_num_in_osds() << " in"; | |
7c673cae FG |
4141 | } |
4142 | ||
3efd9988 | 4143 | bool OSDMap::crush_rule_in_use(int rule_id) const |
7c673cae FG |
4144 | { |
4145 | for (const auto &pool : pools) { | |
3efd9988 | 4146 | if (pool.second.crush_rule == rule_id) |
7c673cae FG |
4147 | return true; |
4148 | } | |
4149 | return false; | |
4150 | } | |
4151 | ||
3efd9988 FG |
4152 | int OSDMap::validate_crush_rules(CrushWrapper *newcrush, |
4153 | ostream *ss) const | |
4154 | { | |
4155 | for (auto& i : pools) { | |
4156 | auto& pool = i.second; | |
4157 | int ruleno = pool.get_crush_rule(); | |
4158 | if (!newcrush->rule_exists(ruleno)) { | |
4159 | *ss << "pool " << i.first << " references crush_rule " << ruleno | |
4160 | << " but it is not present"; | |
4161 | return -EINVAL; | |
4162 | } | |
20effc67 | 4163 | if (newcrush->get_rule_type(ruleno) != (int)pool.get_type()) { |
3efd9988 FG |
4164 | *ss << "pool " << i.first << " type does not match rule " << ruleno; |
4165 | return -EINVAL; | |
4166 | } | |
3efd9988 FG |
4167 | } |
4168 | return 0; | |
4169 | } | |
4170 | ||
224ce89b WB |
4171 | int OSDMap::build_simple_optioned(CephContext *cct, epoch_t e, uuid_d &fsid, |
4172 | int nosd, int pg_bits, int pgp_bits, | |
4173 | bool default_pool) | |
7c673cae | 4174 | { |
224ce89b WB |
4175 | ldout(cct, 10) << "build_simple on " << nosd |
4176 | << " osds" << dendl; | |
7c673cae FG |
4177 | epoch = e; |
4178 | set_fsid(fsid); | |
4179 | created = modified = ceph_clock_now(); | |
4180 | ||
4181 | if (nosd >= 0) { | |
4182 | set_max_osd(nosd); | |
4183 | } else { | |
4184 | // count osds | |
4185 | int maxosd = 0; | |
11fdf7f2 | 4186 | const auto& conf = cct->_conf; |
7c673cae | 4187 | vector<string> sections; |
11fdf7f2 | 4188 | conf.get_all_sections(sections); |
7c673cae FG |
4189 | |
4190 | for (auto §ion : sections) { | |
4191 | if (section.find("osd.") != 0) | |
4192 | continue; | |
4193 | ||
4194 | const char *begin = section.c_str() + 4; | |
4195 | char *end = (char*)begin; | |
4196 | int o = strtol(begin, &end, 10); | |
4197 | if (*end != '\0') | |
4198 | continue; | |
4199 | ||
4200 | if (o > cct->_conf->mon_max_osd) { | |
4201 | lderr(cct) << "[osd." << o << "] in config has id > mon_max_osd " << cct->_conf->mon_max_osd << dendl; | |
4202 | return -ERANGE; | |
4203 | } | |
4204 | ||
4205 | if (o > maxosd) | |
4206 | maxosd = o; | |
4207 | } | |
4208 | ||
4209 | set_max_osd(maxosd + 1); | |
4210 | } | |
4211 | ||
7c673cae FG |
4212 | |
4213 | stringstream ss; | |
4214 | int r; | |
4215 | if (nosd >= 0) | |
4216 | r = build_simple_crush_map(cct, *crush, nosd, &ss); | |
4217 | else | |
4218 | r = build_simple_crush_map_from_conf(cct, *crush, &ss); | |
11fdf7f2 | 4219 | ceph_assert(r == 0); |
7c673cae FG |
4220 | |
4221 | int poolbase = get_max_osd() ? get_max_osd() : 1; | |
4222 | ||
20effc67 | 4223 | const int default_replicated_rule = crush->get_osd_pool_default_crush_replicated_rule(cct); |
11fdf7f2 | 4224 | ceph_assert(default_replicated_rule >= 0); |
7c673cae | 4225 | |
224ce89b WB |
4226 | if (default_pool) { |
4227 | // pgp_num <= pg_num | |
4228 | if (pgp_bits > pg_bits) | |
4229 | pgp_bits = pg_bits; | |
4230 | ||
4231 | vector<string> pool_names; | |
4232 | pool_names.push_back("rbd"); | |
4233 | for (auto &plname : pool_names) { | |
4234 | int64_t pool = ++pool_max; | |
4235 | pools[pool].type = pg_pool_t::TYPE_REPLICATED; | |
4236 | pools[pool].flags = cct->_conf->osd_pool_default_flags; | |
4237 | if (cct->_conf->osd_pool_default_flag_hashpspool) | |
4238 | pools[pool].set_flag(pg_pool_t::FLAG_HASHPSPOOL); | |
4239 | if (cct->_conf->osd_pool_default_flag_nodelete) | |
4240 | pools[pool].set_flag(pg_pool_t::FLAG_NODELETE); | |
4241 | if (cct->_conf->osd_pool_default_flag_nopgchange) | |
4242 | pools[pool].set_flag(pg_pool_t::FLAG_NOPGCHANGE); | |
4243 | if (cct->_conf->osd_pool_default_flag_nosizechange) | |
4244 | pools[pool].set_flag(pg_pool_t::FLAG_NOSIZECHANGE); | |
20effc67 TL |
4245 | if (cct->_conf->osd_pool_default_flag_bulk) |
4246 | pools[pool].set_flag(pg_pool_t::FLAG_BULK); | |
11fdf7f2 TL |
4247 | pools[pool].size = cct->_conf.get_val<uint64_t>("osd_pool_default_size"); |
4248 | pools[pool].min_size = cct->_conf.get_osd_pool_default_min_size( | |
4249 | pools[pool].size); | |
224ce89b WB |
4250 | pools[pool].crush_rule = default_replicated_rule; |
4251 | pools[pool].object_hash = CEPH_STR_HASH_RJENKINS; | |
4252 | pools[pool].set_pg_num(poolbase << pg_bits); | |
4253 | pools[pool].set_pgp_num(poolbase << pgp_bits); | |
11fdf7f2 TL |
4254 | pools[pool].set_pg_num_target(poolbase << pg_bits); |
4255 | pools[pool].set_pgp_num_target(poolbase << pgp_bits); | |
224ce89b | 4256 | pools[pool].last_change = epoch; |
c07f9fc5 FG |
4257 | pools[pool].application_metadata.insert( |
4258 | {pg_pool_t::APPLICATION_NAME_RBD, {}}); | |
9f95a23c TL |
4259 | if (auto m = pg_pool_t::get_pg_autoscale_mode_by_name( |
4260 | cct->_conf.get_val<string>("osd_pool_default_pg_autoscale_mode")); | |
4261 | m != pg_pool_t::pg_autoscale_mode_t::UNKNOWN) { | |
4262 | pools[pool].pg_autoscale_mode = m; | |
4263 | } else { | |
4264 | pools[pool].pg_autoscale_mode = pg_pool_t::pg_autoscale_mode_t::OFF; | |
4265 | } | |
224ce89b WB |
4266 | pool_name[pool] = plname; |
4267 | name_pool[plname] = pool; | |
4268 | } | |
7c673cae FG |
4269 | } |
4270 | ||
7c673cae FG |
4271 | map<string,string> profile_map; |
4272 | r = get_erasure_code_profile_default(cct, profile_map, &ss); | |
4273 | if (r < 0) { | |
4274 | lderr(cct) << ss.str() << dendl; | |
4275 | return r; | |
4276 | } | |
4277 | set_erasure_code_profile("default", profile_map); | |
4278 | return 0; | |
4279 | } | |
4280 | ||
4281 | int OSDMap::get_erasure_code_profile_default(CephContext *cct, | |
4282 | map<string,string> &profile_map, | |
4283 | ostream *ss) | |
4284 | { | |
11fdf7f2 | 4285 | int r = get_json_str_map(cct->_conf.get_val<string>("osd_pool_default_erasure_code_profile"), |
7c673cae FG |
4286 | *ss, |
4287 | &profile_map); | |
4288 | return r; | |
4289 | } | |
4290 | ||
4291 | int OSDMap::_build_crush_types(CrushWrapper& crush) | |
4292 | { | |
4293 | crush.set_type_name(0, "osd"); | |
4294 | crush.set_type_name(1, "host"); | |
4295 | crush.set_type_name(2, "chassis"); | |
4296 | crush.set_type_name(3, "rack"); | |
4297 | crush.set_type_name(4, "row"); | |
4298 | crush.set_type_name(5, "pdu"); | |
4299 | crush.set_type_name(6, "pod"); | |
4300 | crush.set_type_name(7, "room"); | |
4301 | crush.set_type_name(8, "datacenter"); | |
11fdf7f2 TL |
4302 | crush.set_type_name(9, "zone"); |
4303 | crush.set_type_name(10, "region"); | |
4304 | crush.set_type_name(11, "root"); | |
4305 | return 11; | |
7c673cae FG |
4306 | } |
4307 | ||
4308 | int OSDMap::build_simple_crush_map(CephContext *cct, CrushWrapper& crush, | |
4309 | int nosd, ostream *ss) | |
4310 | { | |
4311 | crush.create(); | |
4312 | ||
4313 | // root | |
4314 | int root_type = _build_crush_types(crush); | |
4315 | int rootid; | |
4316 | int r = crush.add_bucket(0, 0, CRUSH_HASH_DEFAULT, | |
4317 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4318 | ceph_assert(r == 0); |
7c673cae FG |
4319 | crush.set_item_name(rootid, "default"); |
4320 | ||
f67539c2 TL |
4321 | map<string,string> loc{ |
4322 | {"host", "localhost"}, | |
4323 | {"rack", "localrack"}, | |
4324 | {"root", "default"} | |
4325 | }; | |
7c673cae | 4326 | for (int o=0; o<nosd; o++) { |
7c673cae FG |
4327 | ldout(cct, 10) << " adding osd." << o << " at " << loc << dendl; |
4328 | char name[32]; | |
4329 | snprintf(name, sizeof(name), "osd.%d", o); | |
4330 | crush.insert_item(cct, o, 1.0, name, loc); | |
4331 | } | |
4332 | ||
31f18b77 | 4333 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4334 | |
4335 | crush.finalize(); | |
4336 | ||
4337 | return 0; | |
4338 | } | |
4339 | ||
4340 | int OSDMap::build_simple_crush_map_from_conf(CephContext *cct, | |
4341 | CrushWrapper& crush, | |
4342 | ostream *ss) | |
4343 | { | |
11fdf7f2 | 4344 | const auto& conf = cct->_conf; |
7c673cae FG |
4345 | |
4346 | crush.create(); | |
4347 | ||
4348 | // root | |
4349 | int root_type = _build_crush_types(crush); | |
4350 | int rootid; | |
4351 | int r = crush.add_bucket(0, 0, | |
4352 | CRUSH_HASH_DEFAULT, | |
4353 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4354 | ceph_assert(r == 0); |
7c673cae FG |
4355 | crush.set_item_name(rootid, "default"); |
4356 | ||
4357 | // add osds | |
4358 | vector<string> sections; | |
11fdf7f2 | 4359 | conf.get_all_sections(sections); |
7c673cae FG |
4360 | |
4361 | for (auto §ion : sections) { | |
4362 | if (section.find("osd.") != 0) | |
4363 | continue; | |
4364 | ||
4365 | const char *begin = section.c_str() + 4; | |
4366 | char *end = (char*)begin; | |
4367 | int o = strtol(begin, &end, 10); | |
4368 | if (*end != '\0') | |
4369 | continue; | |
4370 | ||
4371 | string host, rack, row, room, dc, pool; | |
4372 | vector<string> sectiontmp; | |
4373 | sectiontmp.push_back("osd"); | |
4374 | sectiontmp.push_back(section); | |
11fdf7f2 TL |
4375 | conf.get_val_from_conf_file(sectiontmp, "host", host, false); |
4376 | conf.get_val_from_conf_file(sectiontmp, "rack", rack, false); | |
4377 | conf.get_val_from_conf_file(sectiontmp, "row", row, false); | |
4378 | conf.get_val_from_conf_file(sectiontmp, "room", room, false); | |
4379 | conf.get_val_from_conf_file(sectiontmp, "datacenter", dc, false); | |
4380 | conf.get_val_from_conf_file(sectiontmp, "root", pool, false); | |
7c673cae FG |
4381 | |
4382 | if (host.length() == 0) | |
4383 | host = "unknownhost"; | |
4384 | if (rack.length() == 0) | |
4385 | rack = "unknownrack"; | |
4386 | ||
4387 | map<string,string> loc; | |
4388 | loc["host"] = host; | |
4389 | loc["rack"] = rack; | |
4390 | if (row.size()) | |
4391 | loc["row"] = row; | |
4392 | if (room.size()) | |
4393 | loc["room"] = room; | |
4394 | if (dc.size()) | |
4395 | loc["datacenter"] = dc; | |
4396 | loc["root"] = "default"; | |
4397 | ||
4398 | ldout(cct, 5) << " adding osd." << o << " at " << loc << dendl; | |
4399 | crush.insert_item(cct, o, 1.0, section, loc); | |
4400 | } | |
4401 | ||
31f18b77 | 4402 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4403 | |
4404 | crush.finalize(); | |
4405 | ||
4406 | return 0; | |
4407 | } | |
4408 | ||
4409 | ||
31f18b77 FG |
4410 | int OSDMap::build_simple_crush_rules( |
4411 | CephContext *cct, | |
4412 | CrushWrapper& crush, | |
4413 | const string& root, | |
4414 | ostream *ss) | |
7c673cae | 4415 | { |
20effc67 | 4416 | int crush_rule = crush.get_osd_pool_default_crush_replicated_rule(cct); |
7c673cae FG |
4417 | string failure_domain = |
4418 | crush.get_type_name(cct->_conf->osd_crush_chooseleaf_type); | |
4419 | ||
7c673cae | 4420 | int r; |
31f18b77 | 4421 | r = crush.add_simple_rule_at( |
224ce89b | 4422 | "replicated_rule", root, failure_domain, "", |
31f18b77 FG |
4423 | "firstn", pg_pool_t::TYPE_REPLICATED, |
4424 | crush_rule, ss); | |
7c673cae FG |
4425 | if (r < 0) |
4426 | return r; | |
4427 | // do not add an erasure rule by default or else we will implicitly | |
4428 | // require the crush_v2 feature of clients | |
4429 | return 0; | |
4430 | } | |
4431 | ||
4432 | int OSDMap::summarize_mapping_stats( | |
4433 | OSDMap *newmap, | |
4434 | const set<int64_t> *pools, | |
4435 | std::string *out, | |
4436 | Formatter *f) const | |
4437 | { | |
4438 | set<int64_t> ls; | |
4439 | if (pools) { | |
4440 | ls = *pools; | |
4441 | } else { | |
4442 | for (auto &p : get_pools()) | |
4443 | ls.insert(p.first); | |
4444 | } | |
4445 | ||
4446 | unsigned total_pg = 0; | |
4447 | unsigned moved_pg = 0; | |
4448 | vector<unsigned> base_by_osd(get_max_osd(), 0); | |
4449 | vector<unsigned> new_by_osd(get_max_osd(), 0); | |
4450 | for (int64_t pool_id : ls) { | |
4451 | const pg_pool_t *pi = get_pg_pool(pool_id); | |
31f18b77 FG |
4452 | vector<int> up, up2; |
4453 | int up_primary; | |
7c673cae | 4454 | for (unsigned ps = 0; ps < pi->get_pg_num(); ++ps) { |
11fdf7f2 | 4455 | pg_t pgid(ps, pool_id); |
7c673cae | 4456 | total_pg += pi->get_size(); |
31f18b77 | 4457 | pg_to_up_acting_osds(pgid, &up, &up_primary, nullptr, nullptr); |
7c673cae FG |
4458 | for (int osd : up) { |
4459 | if (osd >= 0 && osd < get_max_osd()) | |
4460 | ++base_by_osd[osd]; | |
4461 | } | |
4462 | if (newmap) { | |
31f18b77 | 4463 | newmap->pg_to_up_acting_osds(pgid, &up2, &up_primary, nullptr, nullptr); |
7c673cae FG |
4464 | for (int osd : up2) { |
4465 | if (osd >= 0 && osd < get_max_osd()) | |
4466 | ++new_by_osd[osd]; | |
4467 | } | |
4468 | if (pi->type == pg_pool_t::TYPE_ERASURE) { | |
4469 | for (unsigned i=0; i<up.size(); ++i) { | |
4470 | if (up[i] != up2[i]) { | |
4471 | ++moved_pg; | |
4472 | } | |
4473 | } | |
4474 | } else if (pi->type == pg_pool_t::TYPE_REPLICATED) { | |
4475 | for (int osd : up) { | |
4476 | if (std::find(up2.begin(), up2.end(), osd) == up2.end()) { | |
4477 | ++moved_pg; | |
4478 | } | |
4479 | } | |
4480 | } else { | |
11fdf7f2 | 4481 | ceph_abort_msg("unhandled pool type"); |
7c673cae FG |
4482 | } |
4483 | } | |
4484 | } | |
4485 | } | |
4486 | ||
4487 | unsigned num_up_in = 0; | |
4488 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4489 | if (is_up(osd) && is_in(osd)) | |
4490 | ++num_up_in; | |
4491 | } | |
4492 | if (!num_up_in) { | |
4493 | return -EINVAL; | |
4494 | } | |
4495 | ||
4496 | float avg_pg = (float)total_pg / (float)num_up_in; | |
4497 | float base_stddev = 0, new_stddev = 0; | |
4498 | int min = -1, max = -1; | |
4499 | unsigned min_base_pg = 0, max_base_pg = 0; | |
4500 | unsigned min_new_pg = 0, max_new_pg = 0; | |
4501 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4502 | if (is_up(osd) && is_in(osd)) { | |
4503 | float base_diff = (float)base_by_osd[osd] - avg_pg; | |
4504 | base_stddev += base_diff * base_diff; | |
4505 | float new_diff = (float)new_by_osd[osd] - avg_pg; | |
4506 | new_stddev += new_diff * new_diff; | |
4507 | if (min < 0 || base_by_osd[osd] < min_base_pg) { | |
4508 | min = osd; | |
4509 | min_base_pg = base_by_osd[osd]; | |
4510 | min_new_pg = new_by_osd[osd]; | |
4511 | } | |
4512 | if (max < 0 || base_by_osd[osd] > max_base_pg) { | |
4513 | max = osd; | |
4514 | max_base_pg = base_by_osd[osd]; | |
4515 | max_new_pg = new_by_osd[osd]; | |
4516 | } | |
4517 | } | |
4518 | } | |
4519 | base_stddev = sqrt(base_stddev / num_up_in); | |
4520 | new_stddev = sqrt(new_stddev / num_up_in); | |
4521 | ||
4522 | float edev = sqrt(avg_pg * (1.0 - (1.0 / (double)num_up_in))); | |
4523 | ||
4524 | ostringstream ss; | |
4525 | if (f) | |
4526 | f->open_object_section("utilization"); | |
4527 | if (newmap) { | |
4528 | if (f) { | |
4529 | f->dump_unsigned("moved_pgs", moved_pg); | |
4530 | f->dump_unsigned("total_pgs", total_pg); | |
4531 | } else { | |
4532 | float percent = 0; | |
4533 | if (total_pg) | |
4534 | percent = (float)moved_pg * 100.0 / (float)total_pg; | |
4535 | ss << "moved " << moved_pg << " / " << total_pg | |
4536 | << " (" << percent << "%)\n"; | |
4537 | } | |
4538 | } | |
4539 | if (f) { | |
4540 | f->dump_float("avg_pgs", avg_pg); | |
4541 | f->dump_float("std_dev", base_stddev); | |
4542 | f->dump_float("expected_baseline_std_dev", edev); | |
4543 | if (newmap) | |
4544 | f->dump_float("new_std_dev", new_stddev); | |
4545 | } else { | |
4546 | ss << "avg " << avg_pg << "\n"; | |
4547 | ss << "stddev " << base_stddev; | |
4548 | if (newmap) | |
4549 | ss << " -> " << new_stddev; | |
4550 | ss << " (expected baseline " << edev << ")\n"; | |
4551 | } | |
4552 | if (min >= 0) { | |
4553 | if (f) { | |
4554 | f->dump_unsigned("min_osd", min); | |
4555 | f->dump_unsigned("min_osd_pgs", min_base_pg); | |
4556 | if (newmap) | |
4557 | f->dump_unsigned("new_min_osd_pgs", min_new_pg); | |
4558 | } else { | |
4559 | ss << "min osd." << min << " with " << min_base_pg; | |
4560 | if (newmap) | |
4561 | ss << " -> " << min_new_pg; | |
4562 | ss << " pgs (" << (float)min_base_pg / avg_pg; | |
4563 | if (newmap) | |
4564 | ss << " -> " << (float)min_new_pg / avg_pg; | |
4565 | ss << " * mean)\n"; | |
4566 | } | |
4567 | } | |
4568 | if (max >= 0) { | |
4569 | if (f) { | |
4570 | f->dump_unsigned("max_osd", max); | |
4571 | f->dump_unsigned("max_osd_pgs", max_base_pg); | |
4572 | if (newmap) | |
4573 | f->dump_unsigned("new_max_osd_pgs", max_new_pg); | |
4574 | } else { | |
4575 | ss << "max osd." << max << " with " << max_base_pg; | |
4576 | if (newmap) | |
4577 | ss << " -> " << max_new_pg; | |
4578 | ss << " pgs (" << (float)max_base_pg / avg_pg; | |
4579 | if (newmap) | |
4580 | ss << " -> " << (float)max_new_pg / avg_pg; | |
4581 | ss << " * mean)\n"; | |
4582 | } | |
4583 | } | |
4584 | if (f) | |
4585 | f->close_section(); | |
4586 | if (out) | |
4587 | *out = ss.str(); | |
4588 | return 0; | |
4589 | } | |
4590 | ||
7c673cae FG |
4591 | bool OSDMap::try_pg_upmap( |
4592 | CephContext *cct, | |
4593 | pg_t pg, ///< pg to potentially remap | |
4594 | const set<int>& overfull, ///< osds we'd want to evacuate | |
4595 | const vector<int>& underfull, ///< osds to move to, in order of preference | |
92f5a8d4 | 4596 | const vector<int>& more_underfull, ///< more osds only slightly underfull |
7c673cae FG |
4597 | vector<int> *orig, |
4598 | vector<int> *out) ///< resulting alternative mapping | |
4599 | { | |
4600 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
4601 | if (!pool) | |
4602 | return false; | |
20effc67 | 4603 | int rule = pool->get_crush_rule(); |
7c673cae FG |
4604 | if (rule < 0) |
4605 | return false; | |
4606 | ||
7c673cae FG |
4607 | // make sure there is something there to remap |
4608 | bool any = false; | |
4609 | for (auto osd : *orig) { | |
4610 | if (overfull.count(osd)) { | |
4611 | any = true; | |
4612 | break; | |
4613 | } | |
4614 | } | |
4615 | if (!any) { | |
4616 | return false; | |
4617 | } | |
4618 | ||
4619 | int r = crush->try_remap_rule( | |
4620 | cct, | |
4621 | rule, | |
4622 | pool->get_size(), | |
4623 | overfull, underfull, | |
92f5a8d4 | 4624 | more_underfull, |
7c673cae FG |
4625 | *orig, |
4626 | out); | |
4627 | if (r < 0) | |
4628 | return false; | |
4629 | if (*out == *orig) | |
4630 | return false; | |
4631 | return true; | |
4632 | } | |
4633 | ||
4634 | int OSDMap::calc_pg_upmaps( | |
4635 | CephContext *cct, | |
92f5a8d4 | 4636 | uint32_t max_deviation, |
7c673cae | 4637 | int max, |
a8e16298 | 4638 | const set<int64_t>& only_pools, |
20effc67 TL |
4639 | OSDMap::Incremental *pending_inc, |
4640 | std::random_device::result_type *p_seed) | |
7c673cae | 4641 | { |
a8e16298 | 4642 | ldout(cct, 10) << __func__ << " pools " << only_pools << dendl; |
20effc67 | 4643 | OSDMap tmp_osd_map; |
92f5a8d4 TL |
4644 | // Can't be less than 1 pg |
4645 | if (max_deviation < 1) | |
4646 | max_deviation = 1; | |
20effc67 | 4647 | tmp_osd_map.deepish_copy_from(*this); |
7c673cae | 4648 | int num_changed = 0; |
a8e16298 TL |
4649 | map<int,set<pg_t>> pgs_by_osd; |
4650 | int total_pgs = 0; | |
4651 | float osd_weight_total = 0; | |
4652 | map<int,float> osd_weight; | |
a8e16298 | 4653 | |
20effc67 TL |
4654 | if (max <= 0) { |
4655 | lderr(cct) << __func__ << " abort due to max <= 0" << dendl; | |
4656 | return 0; | |
a8e16298 | 4657 | } |
20effc67 TL |
4658 | |
4659 | osd_weight_total = build_pool_pgs_info(cct, only_pools, tmp_osd_map, | |
4660 | total_pgs, pgs_by_osd, osd_weight); | |
a8e16298 TL |
4661 | if (osd_weight_total == 0) { |
4662 | lderr(cct) << __func__ << " abort due to osd_weight_total == 0" << dendl; | |
4663 | return 0; | |
4664 | } | |
20effc67 | 4665 | |
a8e16298 TL |
4666 | float pgs_per_weight = total_pgs / osd_weight_total; |
4667 | ldout(cct, 10) << " osd_weight_total " << osd_weight_total << dendl; | |
4668 | ldout(cct, 10) << " pgs_per_weight " << pgs_per_weight << dendl; | |
7c673cae | 4669 | |
a8e16298 TL |
4670 | float stddev = 0; |
4671 | map<int,float> osd_deviation; // osd, deviation(pgs) | |
4672 | multimap<float,int> deviation_osd; // deviation(pgs), osd | |
20effc67 TL |
4673 | float cur_max_deviation = calc_deviations(cct, pgs_by_osd, osd_weight, pgs_per_weight, |
4674 | osd_deviation, deviation_osd, stddev); | |
4675 | ||
92f5a8d4 TL |
4676 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
4677 | if (cur_max_deviation <= max_deviation) { | |
a8e16298 TL |
4678 | ldout(cct, 10) << __func__ << " distribution is almost perfect" |
4679 | << dendl; | |
4680 | return 0; | |
4681 | } | |
20effc67 | 4682 | |
a8e16298 TL |
4683 | bool skip_overfull = false; |
4684 | auto aggressive = | |
11fdf7f2 | 4685 | cct->_conf.get_val<bool>("osd_calc_pg_upmaps_aggressively"); |
a8e16298 | 4686 | auto local_fallback_retries = |
11fdf7f2 | 4687 | cct->_conf.get_val<uint64_t>("osd_calc_pg_upmaps_local_fallback_retries"); |
20effc67 | 4688 | |
a8e16298 | 4689 | while (max--) { |
92f5a8d4 | 4690 | ldout(cct, 30) << "Top of loop #" << max+1 << dendl; |
a8e16298 TL |
4691 | // build overfull and underfull |
4692 | set<int> overfull; | |
92f5a8d4 TL |
4693 | set<int> more_overfull; |
4694 | bool using_more_overfull = false; | |
a8e16298 | 4695 | vector<int> underfull; |
92f5a8d4 | 4696 | vector<int> more_underfull; |
20effc67 TL |
4697 | fill_overfull_underfull(cct, deviation_osd, max_deviation, |
4698 | overfull, more_overfull, | |
4699 | underfull, more_underfull); | |
7c673cae | 4700 | |
92f5a8d4 TL |
4701 | if (underfull.empty() && overfull.empty()) { |
4702 | ldout(cct, 20) << __func__ << " failed to build overfull and underfull" << dendl; | |
7c673cae | 4703 | break; |
a8e16298 | 4704 | } |
92f5a8d4 TL |
4705 | if (overfull.empty() && !underfull.empty()) { |
4706 | ldout(cct, 20) << __func__ << " Using more_overfull since we still have underfull" << dendl; | |
4707 | overfull = more_overfull; | |
4708 | using_more_overfull = true; | |
4709 | } | |
7c673cae | 4710 | |
a8e16298 TL |
4711 | ldout(cct, 10) << " overfull " << overfull |
4712 | << " underfull " << underfull | |
4713 | << dendl; | |
4714 | set<pg_t> to_skip; | |
4715 | uint64_t local_fallback_retried = 0; | |
4716 | ||
4717 | retry: | |
4718 | ||
4719 | set<pg_t> to_unmap; | |
4720 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>> to_upmap; | |
4721 | auto temp_pgs_by_osd = pgs_by_osd; | |
4722 | // always start with fullest, break if we find any changes to make | |
7c673cae | 4723 | for (auto p = deviation_osd.rbegin(); p != deviation_osd.rend(); ++p) { |
92f5a8d4 | 4724 | if (skip_overfull && !underfull.empty()) { |
a8e16298 TL |
4725 | ldout(cct, 10) << " skipping overfull " << dendl; |
4726 | break; // fall through to check underfull | |
4727 | } | |
7c673cae | 4728 | int osd = p->second; |
31f18b77 | 4729 | float deviation = p->first; |
9f95a23c TL |
4730 | if (deviation < 0) { |
4731 | ldout(cct, 10) << " hitting underfull osds now" | |
4732 | << " when trying to remap overfull osds" | |
4733 | << dendl; | |
4734 | break; | |
4735 | } | |
7c673cae | 4736 | float target = osd_weight[osd] * pgs_per_weight; |
92f5a8d4 TL |
4737 | ldout(cct, 10) << " Overfull search osd." << osd |
4738 | << " target " << target | |
4739 | << " deviation " << deviation | |
4740 | << dendl; | |
a8e16298 | 4741 | ceph_assert(target > 0); |
92f5a8d4 | 4742 | if (!using_more_overfull && deviation <= max_deviation) { |
7c673cae | 4743 | ldout(cct, 10) << " osd." << osd |
a8e16298 TL |
4744 | << " target " << target |
4745 | << " deviation " << deviation | |
92f5a8d4 | 4746 | << " < max deviation " << max_deviation |
a8e16298 | 4747 | << dendl; |
7c673cae FG |
4748 | break; |
4749 | } | |
7c673cae | 4750 | |
a8e16298 TL |
4751 | vector<pg_t> pgs; |
4752 | pgs.reserve(pgs_by_osd[osd].size()); | |
4753 | for (auto& pg : pgs_by_osd[osd]) { | |
4754 | if (to_skip.count(pg)) | |
4755 | continue; | |
4756 | pgs.push_back(pg); | |
4757 | } | |
4758 | if (aggressive) { | |
4759 | // shuffle PG list so they all get equal (in)attention | |
20effc67 | 4760 | std::shuffle(pgs.begin(), pgs.end(), get_random_engine(cct, p_seed)); |
a8e16298 | 4761 | } |
7c673cae | 4762 | // look for remaps we can un-remap |
20effc67 TL |
4763 | if (try_drop_remap_overfull(cct, pgs, tmp_osd_map, osd, |
4764 | temp_pgs_by_osd, to_unmap, to_upmap)) | |
4765 | goto test_change; | |
7c673cae | 4766 | |
a8e16298 | 4767 | // try upmap |
7c673cae | 4768 | for (auto pg : pgs) { |
20effc67 TL |
4769 | auto temp_it = tmp_osd_map.pg_upmap.find(pg); |
4770 | if (temp_it != tmp_osd_map.pg_upmap.end()) { | |
a8e16298 TL |
4771 | // leave pg_upmap alone |
4772 | // it must be specified by admin since balancer does not | |
4773 | // support pg_upmap yet | |
4774 | ldout(cct, 10) << " " << pg << " already has pg_upmap " | |
4775 | << temp_it->second << ", skipping" | |
4776 | << dendl; | |
7c673cae FG |
4777 | continue; |
4778 | } | |
20effc67 | 4779 | auto pg_pool_size = tmp_osd_map.get_pg_pool_size(pg); |
a8e16298 TL |
4780 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; |
4781 | set<int> existing; | |
20effc67 TL |
4782 | auto it = tmp_osd_map.pg_upmap_items.find(pg); |
4783 | if (it != tmp_osd_map.pg_upmap_items.end()) { | |
4784 | auto& um_items = it->second; | |
4785 | if (um_items.size() >= (size_t)pg_pool_size) { | |
4786 | ldout(cct, 10) << " " << pg << " already has full-size pg_upmap_items " | |
4787 | << um_items << ", skipping" | |
4788 | << dendl; | |
4789 | continue; | |
4790 | } else { | |
4791 | ldout(cct, 10) << " " << pg << " already has pg_upmap_items " | |
4792 | << um_items | |
4793 | << dendl; | |
4794 | new_upmap_items = um_items; | |
4795 | // build existing too (for dedup) | |
4796 | for (auto [um_from, um_to] : um_items) { | |
4797 | existing.insert(um_from); | |
4798 | existing.insert(um_to); | |
4799 | } | |
4800 | } | |
a8e16298 TL |
4801 | // fall through |
4802 | // to see if we can append more remapping pairs | |
20effc67 | 4803 | } |
a8e16298 | 4804 | ldout(cct, 10) << " trying " << pg << dendl; |
494da23a | 4805 | vector<int> raw, orig, out; |
20effc67 | 4806 | tmp_osd_map.pg_to_raw_upmap(pg, &raw, &orig); // including existing upmaps too |
92f5a8d4 | 4807 | if (!try_pg_upmap(cct, pg, overfull, underfull, more_underfull, &orig, &out)) { |
7c673cae FG |
4808 | continue; |
4809 | } | |
a8e16298 | 4810 | ldout(cct, 10) << " " << pg << " " << orig << " -> " << out << dendl; |
7c673cae FG |
4811 | if (orig.size() != out.size()) { |
4812 | continue; | |
4813 | } | |
a8e16298 | 4814 | ceph_assert(orig != out); |
20effc67 | 4815 | int pos = find_best_remap(cct, orig, out, existing, osd_deviation); |
92f5a8d4 | 4816 | if (pos != -1) { |
a8e16298 TL |
4817 | // append new remapping pairs slowly |
4818 | // This way we can make sure that each tiny change will | |
4819 | // definitely make distribution of PGs converging to | |
4820 | // the perfect status. | |
20effc67 TL |
4821 | add_remap_pair(cct, orig[pos], out[pos], pg, (size_t)pg_pool_size, |
4822 | osd, existing, temp_pgs_by_osd, | |
4823 | new_upmap_items, to_upmap); | |
a8e16298 | 4824 | goto test_change; |
7c673cae | 4825 | } |
a8e16298 TL |
4826 | } |
4827 | } | |
7c673cae | 4828 | |
a8e16298 TL |
4829 | ceph_assert(!(to_unmap.size() || to_upmap.size())); |
4830 | ldout(cct, 10) << " failed to find any changes for overfull osds" | |
4831 | << dendl; | |
20effc67 TL |
4832 | for (auto& [deviation, osd] : deviation_osd) { |
4833 | if (std::find(underfull.begin(), underfull.end(), osd) == | |
a8e16298 TL |
4834 | underfull.end()) |
4835 | break; | |
a8e16298 TL |
4836 | float target = osd_weight[osd] * pgs_per_weight; |
4837 | ceph_assert(target > 0); | |
92f5a8d4 TL |
4838 | if (fabsf(deviation) < max_deviation) { |
4839 | // respect max_deviation too | |
a8e16298 TL |
4840 | ldout(cct, 10) << " osd." << osd |
4841 | << " target " << target | |
4842 | << " deviation " << deviation | |
92f5a8d4 TL |
4843 | << " -> absolute " << fabsf(deviation) |
4844 | << " < max " << max_deviation | |
a8e16298 TL |
4845 | << dendl; |
4846 | break; | |
4847 | } | |
4848 | // look for remaps we can un-remap | |
20effc67 TL |
4849 | candidates_t candidates = build_candidates(cct, tmp_osd_map, to_skip, |
4850 | only_pools, aggressive, p_seed); | |
4851 | if (try_drop_remap_underfull(cct, candidates, osd, temp_pgs_by_osd, | |
4852 | to_unmap, to_upmap)) { | |
4853 | goto test_change; | |
a8e16298 | 4854 | } |
7c673cae | 4855 | } |
a8e16298 TL |
4856 | |
4857 | ceph_assert(!(to_unmap.size() || to_upmap.size())); | |
4858 | ldout(cct, 10) << " failed to find any changes for underfull osds" | |
4859 | << dendl; | |
4860 | if (!aggressive) { | |
4861 | ldout(cct, 10) << " break due to aggressive mode not enabled" << dendl; | |
4862 | break; | |
4863 | } else if (!skip_overfull) { | |
4864 | // safe to quit because below here we know | |
4865 | // we've done checking both overfull and underfull osds.. | |
4866 | ldout(cct, 10) << " break due to not being able to find any" | |
4867 | << " further optimizations" | |
4868 | << dendl; | |
7c673cae FG |
4869 | break; |
4870 | } | |
a8e16298 TL |
4871 | // restart with fullest and do exhaustive searching |
4872 | skip_overfull = false; | |
4873 | continue; | |
4874 | ||
4875 | test_change: | |
4876 | ||
4877 | // test change, apply if change is good | |
4878 | ceph_assert(to_unmap.size() || to_upmap.size()); | |
4879 | float new_stddev = 0; | |
4880 | map<int,float> temp_osd_deviation; | |
4881 | multimap<float,int> temp_deviation_osd; | |
20effc67 TL |
4882 | float cur_max_deviation = calc_deviations(cct, temp_pgs_by_osd, osd_weight, |
4883 | pgs_per_weight, temp_osd_deviation, | |
4884 | temp_deviation_osd, new_stddev); | |
a8e16298 TL |
4885 | ldout(cct, 10) << " stddev " << stddev << " -> " << new_stddev << dendl; |
4886 | if (new_stddev >= stddev) { | |
4887 | if (!aggressive) { | |
4888 | ldout(cct, 10) << " break because stddev is not decreasing" | |
4889 | << " and aggressive mode is not enabled" | |
4890 | << dendl; | |
4891 | break; | |
4892 | } | |
4893 | local_fallback_retried++; | |
4894 | if (local_fallback_retried >= local_fallback_retries) { | |
4895 | // does not make progress | |
4896 | // flip *skip_overfull* so both overfull and underfull | |
4897 | // get equal (in)attention | |
4898 | skip_overfull = !skip_overfull; | |
4899 | ldout(cct, 10) << " hit local_fallback_retries " | |
4900 | << local_fallback_retries | |
4901 | << dendl; | |
4902 | continue; | |
4903 | } | |
4904 | for (auto& i : to_unmap) | |
4905 | to_skip.insert(i); | |
4906 | for (auto& i : to_upmap) | |
4907 | to_skip.insert(i.first); | |
4908 | ldout(cct, 20) << " local_fallback_retried " << local_fallback_retried | |
4909 | << " to_skip " << to_skip | |
4910 | << dendl; | |
4911 | goto retry; | |
4912 | } | |
4913 | ||
4914 | // ready to go | |
4915 | ceph_assert(new_stddev < stddev); | |
4916 | stddev = new_stddev; | |
4917 | pgs_by_osd = temp_pgs_by_osd; | |
4918 | osd_deviation = temp_osd_deviation; | |
4919 | deviation_osd = temp_deviation_osd; | |
20effc67 TL |
4920 | |
4921 | num_changed += pack_upmap_results(cct, to_unmap, to_upmap, tmp_osd_map, pending_inc); | |
4922 | ||
92f5a8d4 TL |
4923 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
4924 | if (cur_max_deviation <= max_deviation) { | |
4925 | ldout(cct, 10) << __func__ << " Optimization plan is almost perfect" | |
4926 | << dendl; | |
4927 | break; | |
4928 | } | |
7c673cae | 4929 | } |
a8e16298 | 4930 | ldout(cct, 10) << " num_changed = " << num_changed << dendl; |
7c673cae FG |
4931 | return num_changed; |
4932 | } | |
31f18b77 | 4933 | |
20effc67 TL |
4934 | float OSDMap::build_pool_pgs_info ( |
4935 | CephContext *cct, | |
4936 | const std::set<int64_t>& only_pools, ///< [optional] restrict to pool | |
4937 | const OSDMap& tmp_osd_map, | |
4938 | int& total_pgs, | |
4939 | map<int,set<pg_t>>& pgs_by_osd, | |
4940 | map<int,float>& osd_weight) | |
4941 | { | |
4942 | // | |
4943 | // This function builds some data structures that are used by calc_pg_upmaps. | |
4944 | // Specifically it builds pgs_by_osd and osd_weight maps, updates total_pgs | |
4945 | // and returns the osd_weight_total | |
4946 | // | |
4947 | float osd_weight_total = 0.0; | |
4948 | for (auto& [pid, pdata] : pools) { | |
4949 | if (!only_pools.empty() && !only_pools.count(pid)) | |
4950 | continue; | |
4951 | for (unsigned ps = 0; ps < pdata.get_pg_num(); ++ps) { | |
4952 | pg_t pg(ps, pid); | |
4953 | vector<int> up; | |
4954 | tmp_osd_map.pg_to_up_acting_osds(pg, &up, nullptr, nullptr, nullptr); | |
4955 | ldout(cct, 20) << __func__ << " " << pg << " up " << up << dendl; | |
4956 | for (auto osd : up) { | |
4957 | if (osd != CRUSH_ITEM_NONE) | |
4958 | pgs_by_osd[osd].insert(pg); | |
4959 | } | |
4960 | } | |
4961 | total_pgs += pdata.get_size() * pdata.get_pg_num(); | |
4962 | ||
4963 | map<int,float> pmap; | |
4964 | int ruleno = pdata.get_crush_rule(); | |
4965 | tmp_osd_map.crush->get_rule_weight_osd_map(ruleno, &pmap); | |
4966 | ldout(cct,20) << __func__ << " pool " << pid | |
4967 | << " ruleno " << ruleno | |
4968 | << " weight-map " << pmap | |
4969 | << dendl; | |
4970 | for (auto [oid, oweight] : pmap) { | |
4971 | auto adjusted_weight = tmp_osd_map.get_weightf(oid) * oweight; | |
4972 | if (adjusted_weight == 0) { | |
4973 | continue; | |
4974 | } | |
4975 | osd_weight[oid] += adjusted_weight; | |
4976 | osd_weight_total += adjusted_weight; | |
4977 | } | |
4978 | } | |
4979 | for (auto& [oid, oweight] : osd_weight) { | |
4980 | int pgs = 0; | |
4981 | auto p = pgs_by_osd.find(oid); | |
4982 | if (p != pgs_by_osd.end()) | |
4983 | pgs = p->second.size(); | |
4984 | else | |
4985 | pgs_by_osd.emplace(oid, set<pg_t>()); | |
4986 | ldout(cct, 20) << " osd." << oid << " weight " << oweight | |
4987 | << " pgs " << pgs << dendl; | |
4988 | } | |
4989 | return osd_weight_total; | |
4990 | ||
4991 | } // return total weight of all OSDs | |
4992 | ||
4993 | float OSDMap::calc_deviations ( | |
4994 | CephContext *cct, | |
4995 | const map<int,set<pg_t>>& pgs_by_osd, | |
4996 | const map<int,float>& osd_weight, | |
4997 | float pgs_per_weight, | |
4998 | map<int,float>& osd_deviation, | |
4999 | multimap<float,int>& deviation_osd, | |
5000 | float& stddev) // return current max deviation | |
5001 | { | |
5002 | // | |
5003 | // This function calculates the 2 maps osd_deviation and deviation_osd which | |
5004 | // hold the deviation between the current number of PGs which map to an OSD | |
5005 | // and the optimal number. Ot also calculates the stddev of the deviations and | |
5006 | // returns the current max deviation. | |
5007 | // NOTE - the calculation is not exactly stddev it is actually sttdev^2 but as | |
5008 | // long as it is monotonic with stddev (and it is), it is sufficient for | |
5009 | // the balancer code. | |
5010 | // | |
5011 | float cur_max_deviation = 0.0; | |
5012 | stddev = 0.0; | |
5013 | for (auto& [oid, opgs] : pgs_by_osd) { | |
5014 | // make sure osd is still there (belongs to this crush-tree) | |
5015 | ceph_assert(osd_weight.count(oid)); | |
5016 | float target = osd_weight.at(oid) * pgs_per_weight; | |
5017 | float deviation = (float)opgs.size() - target; | |
5018 | ldout(cct, 20) << " osd." << oid | |
5019 | << "\tpgs " << opgs.size() | |
5020 | << "\ttarget " << target | |
5021 | << "\tdeviation " << deviation | |
5022 | << dendl; | |
5023 | osd_deviation[oid] = deviation; | |
5024 | deviation_osd.insert(make_pair(deviation, oid)); | |
5025 | stddev += deviation * deviation; | |
5026 | if (fabsf(deviation) > cur_max_deviation) | |
5027 | cur_max_deviation = fabsf(deviation); | |
5028 | } | |
5029 | return cur_max_deviation; | |
5030 | } | |
5031 | ||
5032 | void OSDMap::fill_overfull_underfull ( | |
5033 | CephContext *cct, | |
5034 | const std::multimap<float,int>& deviation_osd, | |
5035 | int max_deviation, | |
5036 | std::set<int>& overfull, | |
5037 | std::set<int>& more_overfull, | |
5038 | std::vector<int>& underfull, | |
5039 | std::vector<int>& more_underfull) | |
5040 | { | |
5041 | // | |
5042 | // This function just fills the overfull and underfull data structures for the | |
5043 | // use of calc_pg_upmaps | |
5044 | // | |
5045 | for (auto i = deviation_osd.rbegin(); i != deviation_osd.rend(); i++) { | |
5046 | auto& odev = i->first; | |
5047 | auto& oid = i->second; | |
5048 | ldout(cct, 30) << " check " << odev << " <= " << max_deviation << dendl; | |
5049 | if (odev <= 0) | |
5050 | break; | |
5051 | if (odev > max_deviation) { | |
5052 | ldout(cct, 30) << " add overfull osd." << oid << dendl; | |
5053 | overfull.insert(oid); | |
5054 | } else { | |
5055 | more_overfull.insert(oid); | |
5056 | } | |
5057 | } | |
5058 | ||
5059 | for (auto i = deviation_osd.begin(); i != deviation_osd.end(); i++) { | |
5060 | auto& odev = i->first; | |
5061 | auto& oid = i->second; | |
5062 | ldout(cct, 30) << " check " << odev << " >= " << -(int)max_deviation << dendl; | |
5063 | if (odev >= 0) | |
5064 | break; | |
5065 | if (odev < -(int)max_deviation) { | |
5066 | ldout(cct, 30) << " add underfull osd." << oid << dendl; | |
5067 | underfull.push_back(oid); | |
5068 | } else { | |
5069 | more_underfull.push_back(oid); | |
5070 | } | |
5071 | } | |
5072 | } | |
5073 | ||
5074 | int OSDMap::pack_upmap_results( | |
5075 | CephContext *cct, | |
5076 | const std::set<pg_t>& to_unmap, | |
5077 | const std::map<pg_t, mempool::osdmap::vector<std::pair<int, int>>>& to_upmap, | |
5078 | OSDMap& tmp_osd_map, | |
5079 | OSDMap::Incremental *pending_inc) | |
5080 | { | |
5081 | // | |
5082 | // This function takes the input from the local variables to_unmap and to_upmap | |
5083 | // and updates tmp_osd_map (so that another iteration can run) and pending_inc | |
5084 | // (so that the results are visible outside calc_pg_upmaps) | |
5085 | // | |
5086 | int num_changed = 0; | |
5087 | for (auto& i : to_unmap) { | |
5088 | ldout(cct, 10) << " unmap pg " << i << dendl; | |
5089 | ceph_assert(tmp_osd_map.pg_upmap_items.count(i)); | |
5090 | tmp_osd_map.pg_upmap_items.erase(i); | |
5091 | pending_inc->old_pg_upmap_items.insert(i); | |
5092 | ++num_changed; | |
5093 | } | |
5094 | for (auto& [pg, um_items] : to_upmap) { | |
5095 | ldout(cct, 10) << " upmap pg " << pg | |
5096 | << " new pg_upmap_items " << um_items | |
5097 | << dendl; | |
5098 | tmp_osd_map.pg_upmap_items[pg] = um_items; | |
5099 | pending_inc->new_pg_upmap_items[pg] = um_items; | |
5100 | ++num_changed; | |
5101 | } | |
5102 | ||
5103 | return num_changed; | |
5104 | } | |
5105 | ||
5106 | std::default_random_engine OSDMap::get_random_engine( | |
5107 | CephContext *cct, | |
5108 | std::random_device::result_type *p_seed) | |
5109 | { | |
5110 | // | |
5111 | // This function creates a random_engine to be used for shuffling. | |
5112 | // When p_seed == nullptr it generates random engine with a seed from /dev/random | |
5113 | // when p_seed is not null, it uses (*p_seed + seed_set) as the seed and | |
5114 | // increments seed_set. This is used in order to craete regression test without | |
5115 | // random effect on the results. | |
5116 | // | |
5117 | static std::random_device::result_type seed_set = 0; | |
5118 | std::random_device::result_type seed; | |
5119 | if (p_seed == nullptr) { | |
5120 | std::random_device rd; | |
5121 | seed = rd(); | |
5122 | } | |
5123 | else { | |
5124 | seed = *p_seed + seed_set; | |
5125 | ldout(cct, 30) << " Starting random engine with seed " | |
5126 | << seed << dendl; | |
5127 | seed_set++; | |
5128 | } | |
5129 | return std::default_random_engine{seed}; | |
5130 | } | |
5131 | ||
5132 | bool OSDMap::try_drop_remap_overfull( | |
5133 | CephContext *cct, | |
5134 | const std::vector<pg_t>& pgs, | |
5135 | const OSDMap& tmp_osd_map, | |
5136 | int osd, | |
5137 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5138 | set<pg_t>& to_unmap, | |
5139 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5140 | { | |
5141 | // | |
5142 | // This function tries to drop existimg upmap items which map data to overfull | |
5143 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5144 | // if it found an item that can be dropped, false if not. | |
5145 | // | |
5146 | for (auto pg : pgs) { | |
5147 | auto p = tmp_osd_map.pg_upmap_items.find(pg); | |
5148 | if (p == tmp_osd_map.pg_upmap_items.end()) | |
5149 | continue; | |
5150 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5151 | auto& pg_upmap_items = p->second; | |
5152 | for (auto um_pair : pg_upmap_items) { | |
5153 | auto& um_from = um_pair.first; | |
5154 | auto& um_to = um_pair.second; | |
5155 | if (um_to == osd) { | |
5156 | ldout(cct, 10) << " will try dropping existing" | |
5157 | << " remapping pair " | |
5158 | << um_from << " -> " << um_to | |
5159 | << " which remapped " << pg | |
5160 | << " into overfull osd." << osd | |
5161 | << dendl; | |
5162 | temp_pgs_by_osd[um_to].erase(pg); | |
5163 | temp_pgs_by_osd[um_from].insert(pg); | |
5164 | } else { | |
5165 | new_upmap_items.push_back(um_pair); | |
5166 | } | |
5167 | } | |
5168 | if (new_upmap_items.empty()) { | |
5169 | // drop whole item | |
5170 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5171 | << " remapped " << pg << " into overfull osd." << osd | |
5172 | << ", will try cancelling it entirely" | |
5173 | << dendl; | |
5174 | to_unmap.insert(pg); | |
5175 | return true; | |
5176 | } else if (new_upmap_items.size() != pg_upmap_items.size()) { | |
5177 | // drop single remapping pair, updating | |
5178 | ceph_assert(new_upmap_items.size() < pg_upmap_items.size()); | |
5179 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5180 | << " remapped " << pg << " into overfull osd." << osd | |
5181 | << ", new_pg_upmap_items now " << new_upmap_items | |
5182 | << dendl; | |
5183 | to_upmap[pg] = new_upmap_items; | |
5184 | return true; | |
5185 | } | |
5186 | } | |
5187 | return false; | |
5188 | } | |
5189 | ||
5190 | bool OSDMap::try_drop_remap_underfull( | |
5191 | CephContext *cct, | |
5192 | const candidates_t& candidates, | |
5193 | int osd, | |
5194 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5195 | set<pg_t>& to_unmap, | |
5196 | map<pg_t, mempool::osdmap::vector<std::pair<int32_t,int32_t>>>& to_upmap) | |
5197 | { | |
5198 | // | |
5199 | // This function tries to drop existimg upmap items which map data from underfull | |
5200 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5201 | // if it found an item that can be dropped, false if not. | |
5202 | // | |
5203 | for (auto& [pg, um_pairs] : candidates) { | |
5204 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5205 | for (auto& ump : um_pairs) { | |
5206 | auto& um_from = ump.first; | |
5207 | auto& um_to = ump.second; | |
5208 | if (um_from == osd) { | |
5209 | ldout(cct, 10) << " will try dropping existing" | |
5210 | << " remapping pair " | |
5211 | << um_from << " -> " << um_to | |
5212 | << " which remapped " << pg | |
5213 | << " out from underfull osd." << osd | |
5214 | << dendl; | |
5215 | temp_pgs_by_osd[um_to].erase(pg); | |
5216 | temp_pgs_by_osd[um_from].insert(pg); | |
5217 | } else { | |
5218 | new_upmap_items.push_back(ump); | |
5219 | } | |
5220 | } | |
5221 | if (new_upmap_items.empty()) { | |
5222 | // drop whole item | |
5223 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5224 | << " remapped " << pg | |
5225 | << " out from underfull osd." << osd | |
5226 | << ", will try cancelling it entirely" | |
5227 | << dendl; | |
5228 | to_unmap.insert(pg); | |
5229 | return true; | |
5230 | } else if (new_upmap_items.size() != um_pairs.size()) { | |
5231 | // drop single remapping pair, updating | |
5232 | ceph_assert(new_upmap_items.size() < um_pairs.size()); | |
5233 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5234 | << " remapped " << pg | |
5235 | << " out from underfull osd." << osd | |
5236 | << ", new_pg_upmap_items now " << new_upmap_items | |
5237 | << dendl; | |
5238 | to_upmap[pg] = new_upmap_items; | |
5239 | return true; | |
5240 | } | |
5241 | } | |
5242 | return false; | |
5243 | } | |
5244 | ||
5245 | void OSDMap::add_remap_pair( | |
5246 | CephContext *cct, | |
5247 | int orig, | |
5248 | int out, | |
5249 | pg_t pg, | |
5250 | size_t pg_pool_size, | |
5251 | int osd, | |
5252 | set<int>& existing, | |
5253 | map<int,set<pg_t>>& temp_pgs_by_osd, | |
5254 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items, | |
5255 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5256 | { | |
5257 | // | |
5258 | // add a single remap pair (in pg <pg> remap osd from <orig> to <out>) to all | |
5259 | // the relevant data structures | |
5260 | // | |
5261 | ldout(cct, 10) << " will try adding new remapping pair " | |
5262 | << orig << " -> " << out << " for " << pg | |
5263 | << (orig != osd ? " NOT selected osd" : "") | |
5264 | << dendl; | |
5265 | existing.insert(orig); | |
5266 | existing.insert(out); | |
5267 | temp_pgs_by_osd[orig].erase(pg); | |
5268 | temp_pgs_by_osd[out].insert(pg); | |
5269 | ceph_assert(new_upmap_items.size() < pg_pool_size); | |
5270 | new_upmap_items.push_back(make_pair(orig, out)); | |
5271 | // append new remapping pairs slowly | |
5272 | // This way we can make sure that each tiny change will | |
5273 | // definitely make distribution of PGs converging to | |
5274 | // the perfect status. | |
5275 | to_upmap[pg] = new_upmap_items; | |
5276 | ||
5277 | } | |
5278 | ||
5279 | int OSDMap::find_best_remap ( | |
5280 | CephContext *cct, | |
5281 | const vector<int>& orig, | |
5282 | const vector<int>& out, | |
5283 | const set<int>& existing, | |
5284 | const map<int,float> osd_deviation) | |
5285 | { | |
5286 | // | |
5287 | // Find the best remap from the suggestions in orig and out - the best remap | |
5288 | // is the one which maps from the OSD with the largest deviatoion (from the | |
5289 | // OSDs which are part of orig) | |
5290 | // | |
5291 | int best_pos = -1; | |
5292 | float max_dev = 0; | |
5293 | for (unsigned i = 0; i < out.size(); ++i) { | |
5294 | if (orig[i] == out[i]) | |
5295 | continue; // skip invalid remappings | |
5296 | if (existing.count(orig[i]) || existing.count(out[i])) | |
5297 | continue; // we want new remappings only! | |
5298 | if (osd_deviation.at(orig[i]) > max_dev) { | |
5299 | max_dev = osd_deviation.at(orig[i]); | |
5300 | best_pos = i; | |
5301 | ldout(cct, 30) << "Max osd." << orig[i] << " pos " << i << " dev " << osd_deviation.at(orig[i]) << dendl; | |
5302 | } | |
5303 | } | |
5304 | return best_pos; | |
5305 | } | |
5306 | ||
5307 | OSDMap::candidates_t OSDMap::build_candidates( | |
5308 | CephContext *cct, | |
5309 | const OSDMap& tmp_osd_map, | |
5310 | const set<pg_t> to_skip, | |
5311 | const set<int64_t>& only_pools, | |
5312 | bool aggressive, | |
5313 | std::random_device::result_type *p_seed) | |
5314 | { | |
5315 | // | |
5316 | // build the candidates data structure | |
5317 | // | |
5318 | candidates_t candidates; | |
5319 | candidates.reserve(tmp_osd_map.pg_upmap_items.size()); | |
5320 | for (auto& [pg, um_pair] : tmp_osd_map.pg_upmap_items) { | |
5321 | if (to_skip.count(pg)) | |
5322 | continue; | |
5323 | if (!only_pools.empty() && !only_pools.count(pg.pool())) | |
5324 | continue; | |
5325 | candidates.push_back(make_pair(pg, um_pair)); | |
5326 | } | |
5327 | if (aggressive) { | |
5328 | // shuffle candidates so they all get equal (in)attention | |
5329 | std::shuffle(candidates.begin(), candidates.end(), get_random_engine(cct, p_seed)); | |
5330 | } | |
5331 | return candidates; | |
5332 | } | |
5333 | ||
31f18b77 FG |
5334 | int OSDMap::get_osds_by_bucket_name(const string &name, set<int> *osds) const |
5335 | { | |
5336 | return crush->get_leaves(name, osds); | |
5337 | } | |
5338 | ||
3efd9988 FG |
5339 | // get pools whose crush rules might reference the given osd |
5340 | void OSDMap::get_pool_ids_by_osd(CephContext *cct, | |
5341 | int osd, | |
5342 | set<int64_t> *pool_ids) const | |
5343 | { | |
11fdf7f2 | 5344 | ceph_assert(pool_ids); |
3efd9988 FG |
5345 | set<int> raw_rules; |
5346 | int r = crush->get_rules_by_osd(osd, &raw_rules); | |
5347 | if (r < 0) { | |
5348 | lderr(cct) << __func__ << " get_rules_by_osd failed: " << cpp_strerror(r) | |
5349 | << dendl; | |
11fdf7f2 | 5350 | ceph_assert(r >= 0); |
3efd9988 FG |
5351 | } |
5352 | set<int> rules; | |
5353 | for (auto &i: raw_rules) { | |
5354 | // exclude any dead rule | |
5355 | if (crush_rule_in_use(i)) { | |
5356 | rules.insert(i); | |
5357 | } | |
5358 | } | |
5359 | for (auto &r: rules) { | |
5360 | get_pool_ids_by_rule(r, pool_ids); | |
5361 | } | |
5362 | } | |
5363 | ||
31f18b77 FG |
5364 | template <typename F> |
5365 | class OSDUtilizationDumper : public CrushTreeDumper::Dumper<F> { | |
5366 | public: | |
5367 | typedef CrushTreeDumper::Dumper<F> Parent; | |
5368 | ||
5369 | OSDUtilizationDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
11fdf7f2 | 5370 | const PGMap& pgmap_, bool tree_, |
9f95a23c | 5371 | const string& filter) : |
c07f9fc5 | 5372 | Parent(crush, osdmap_->get_pool_names()), |
31f18b77 | 5373 | osdmap(osdmap_), |
11fdf7f2 | 5374 | pgmap(pgmap_), |
31f18b77 | 5375 | tree(tree_), |
31f18b77 FG |
5376 | min_var(-1), |
5377 | max_var(-1), | |
5378 | stddev(0), | |
5379 | sum(0) { | |
9f95a23c TL |
5380 | if (osdmap->crush->name_exists(filter)) { |
5381 | // filter by crush node | |
5382 | auto item_id = osdmap->crush->get_item_id(filter); | |
11fdf7f2 TL |
5383 | allowed.insert(item_id); |
5384 | osdmap->crush->get_all_children(item_id, &allowed); | |
9f95a23c TL |
5385 | } else if (osdmap->crush->class_exists(filter)) { |
5386 | // filter by device class | |
5387 | class_id = osdmap->crush->get_class_id(filter); | |
5388 | } else if (auto pool_id = osdmap->lookup_pg_pool_name(filter); | |
5389 | pool_id >= 0) { | |
5390 | // filter by pool | |
5391 | auto crush_rule = osdmap->get_pool_crush_rule(pool_id); | |
5392 | set<int> roots; | |
5393 | osdmap->crush->find_takes_by_rule(crush_rule, &roots); | |
5394 | allowed = roots; | |
5395 | for (auto r : roots) | |
5396 | osdmap->crush->get_all_children(r, &allowed); | |
11fdf7f2 TL |
5397 | } |
5398 | average_util = average_utilization(); | |
31f18b77 FG |
5399 | } |
5400 | ||
5401 | protected: | |
11fdf7f2 TL |
5402 | |
5403 | bool should_dump(int id) const { | |
5404 | if (!allowed.empty() && !allowed.count(id)) // filter by name | |
5405 | return false; | |
9f95a23c TL |
5406 | if (id >= 0 && class_id >= 0) { |
5407 | auto item_class_id = osdmap->crush->get_item_class_id(id); | |
5408 | if (item_class_id < 0 || // not bound to a class yet | |
5409 | item_class_id != class_id) // or already bound to a different class | |
11fdf7f2 TL |
5410 | return false; |
5411 | } | |
5412 | return true; | |
5413 | } | |
5414 | ||
5415 | set<int> get_dumped_osds() { | |
9f95a23c | 5416 | if (allowed.empty() && class_id < 0) { |
11fdf7f2 TL |
5417 | // old way, all |
5418 | return {}; | |
5419 | } | |
5420 | return dumped_osds; | |
5421 | } | |
5422 | ||
31f18b77 FG |
5423 | void dump_stray(F *f) { |
5424 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
5425 | if (osdmap->exists(i) && !this->is_touched(i)) | |
c07f9fc5 | 5426 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
31f18b77 FG |
5427 | } |
5428 | } | |
5429 | ||
5430 | void dump_item(const CrushTreeDumper::Item &qi, F *f) override { | |
f67539c2 | 5431 | if (!tree && (qi.is_bucket() || dumped_osds.count(qi.id))) |
31f18b77 | 5432 | return; |
11fdf7f2 TL |
5433 | if (!should_dump(qi.id)) |
5434 | return; | |
31f18b77 | 5435 | |
11fdf7f2 TL |
5436 | if (!qi.is_bucket()) |
5437 | dumped_osds.insert(qi.id); | |
31f18b77 | 5438 | float reweight = qi.is_bucket() ? -1 : osdmap->get_weightf(qi.id); |
11fdf7f2 TL |
5439 | int64_t kb = 0, kb_used = 0, kb_used_data = 0, kb_used_omap = 0, |
5440 | kb_used_meta = 0, kb_avail = 0; | |
31f18b77 | 5441 | double util = 0; |
11fdf7f2 TL |
5442 | if (get_bucket_utilization(qi.id, &kb, &kb_used, &kb_used_data, |
5443 | &kb_used_omap, &kb_used_meta, &kb_avail)) | |
31f18b77 FG |
5444 | if (kb_used && kb) |
5445 | util = 100.0 * (double)kb_used / (double)kb; | |
5446 | ||
5447 | double var = 1.0; | |
5448 | if (average_util) | |
5449 | var = util / average_util; | |
5450 | ||
11fdf7f2 | 5451 | size_t num_pgs = qi.is_bucket() ? 0 : pgmap.get_num_pg_by_osd(qi.id); |
31f18b77 | 5452 | |
11fdf7f2 TL |
5453 | dump_item(qi, reweight, kb, kb_used, |
5454 | kb_used_data, kb_used_omap, kb_used_meta, | |
5455 | kb_avail, util, var, num_pgs, f); | |
31f18b77 FG |
5456 | |
5457 | if (!qi.is_bucket() && reweight > 0) { | |
5458 | if (min_var < 0 || var < min_var) | |
5459 | min_var = var; | |
5460 | if (max_var < 0 || var > max_var) | |
5461 | max_var = var; | |
5462 | ||
5463 | double dev = util - average_util; | |
5464 | dev *= dev; | |
5465 | stddev += reweight * dev; | |
5466 | sum += reweight; | |
5467 | } | |
5468 | } | |
5469 | ||
5470 | virtual void dump_item(const CrushTreeDumper::Item &qi, | |
5471 | float &reweight, | |
5472 | int64_t kb, | |
5473 | int64_t kb_used, | |
11fdf7f2 TL |
5474 | int64_t kb_used_data, |
5475 | int64_t kb_used_omap, | |
5476 | int64_t kb_used_meta, | |
31f18b77 FG |
5477 | int64_t kb_avail, |
5478 | double& util, | |
5479 | double& var, | |
5480 | const size_t num_pgs, | |
5481 | F *f) = 0; | |
5482 | ||
5483 | double dev() { | |
5484 | return sum > 0 ? sqrt(stddev / sum) : 0; | |
5485 | } | |
5486 | ||
5487 | double average_utilization() { | |
5488 | int64_t kb = 0, kb_used = 0; | |
5489 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
11fdf7f2 TL |
5490 | if (!osdmap->exists(i) || |
5491 | osdmap->get_weight(i) == 0 || | |
5492 | !should_dump(i)) | |
31f18b77 | 5493 | continue; |
11fdf7f2 TL |
5494 | int64_t kb_i, kb_used_i, kb_used_data_i, kb_used_omap_i, kb_used_meta_i, |
5495 | kb_avail_i; | |
5496 | if (get_osd_utilization(i, &kb_i, &kb_used_i, &kb_used_data_i, | |
5497 | &kb_used_omap_i, &kb_used_meta_i, &kb_avail_i)) { | |
31f18b77 FG |
5498 | kb += kb_i; |
5499 | kb_used += kb_used_i; | |
5500 | } | |
5501 | } | |
5502 | return kb > 0 ? 100.0 * (double)kb_used / (double)kb : 0; | |
5503 | } | |
5504 | ||
5505 | bool get_osd_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5506 | int64_t* kb_used_data, |
5507 | int64_t* kb_used_omap, | |
5508 | int64_t* kb_used_meta, | |
31f18b77 | 5509 | int64_t* kb_avail) const { |
11fdf7f2 | 5510 | const osd_stat_t *p = pgmap.get_osd_stat(id); |
31f18b77 | 5511 | if (!p) return false; |
11fdf7f2 TL |
5512 | *kb = p->statfs.kb(); |
5513 | *kb_used = p->statfs.kb_used_raw(); | |
5514 | *kb_used_data = p->statfs.kb_used_data(); | |
5515 | *kb_used_omap = p->statfs.kb_used_omap(); | |
5516 | *kb_used_meta = p->statfs.kb_used_internal_metadata(); | |
5517 | *kb_avail = p->statfs.kb_avail(); | |
5518 | ||
f67539c2 | 5519 | return true; |
31f18b77 FG |
5520 | } |
5521 | ||
5522 | bool get_bucket_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5523 | int64_t* kb_used_data, |
5524 | int64_t* kb_used_omap, | |
5525 | int64_t* kb_used_meta, | |
31f18b77 FG |
5526 | int64_t* kb_avail) const { |
5527 | if (id >= 0) { | |
11fdf7f2 | 5528 | if (osdmap->is_out(id) || !should_dump(id)) { |
31f18b77 FG |
5529 | *kb = 0; |
5530 | *kb_used = 0; | |
11fdf7f2 TL |
5531 | *kb_used_data = 0; |
5532 | *kb_used_omap = 0; | |
5533 | *kb_used_meta = 0; | |
31f18b77 FG |
5534 | *kb_avail = 0; |
5535 | return true; | |
5536 | } | |
11fdf7f2 TL |
5537 | return get_osd_utilization(id, kb, kb_used, kb_used_data, |
5538 | kb_used_omap, kb_used_meta, kb_avail); | |
31f18b77 FG |
5539 | } |
5540 | ||
5541 | *kb = 0; | |
5542 | *kb_used = 0; | |
11fdf7f2 TL |
5543 | *kb_used_data = 0; |
5544 | *kb_used_omap = 0; | |
5545 | *kb_used_meta = 0; | |
31f18b77 FG |
5546 | *kb_avail = 0; |
5547 | ||
5548 | for (int k = osdmap->crush->get_bucket_size(id) - 1; k >= 0; k--) { | |
5549 | int item = osdmap->crush->get_bucket_item(id, k); | |
11fdf7f2 TL |
5550 | int64_t kb_i = 0, kb_used_i = 0, kb_used_data_i = 0, |
5551 | kb_used_omap_i = 0, kb_used_meta_i = 0, kb_avail_i = 0; | |
5552 | if (!get_bucket_utilization(item, &kb_i, &kb_used_i, | |
5553 | &kb_used_data_i, &kb_used_omap_i, | |
5554 | &kb_used_meta_i, &kb_avail_i)) | |
31f18b77 FG |
5555 | return false; |
5556 | *kb += kb_i; | |
5557 | *kb_used += kb_used_i; | |
11fdf7f2 TL |
5558 | *kb_used_data += kb_used_data_i; |
5559 | *kb_used_omap += kb_used_omap_i; | |
5560 | *kb_used_meta += kb_used_meta_i; | |
31f18b77 FG |
5561 | *kb_avail += kb_avail_i; |
5562 | } | |
f67539c2 | 5563 | return true; |
31f18b77 FG |
5564 | } |
5565 | ||
5566 | protected: | |
5567 | const OSDMap *osdmap; | |
11fdf7f2 | 5568 | const PGMap& pgmap; |
31f18b77 FG |
5569 | bool tree; |
5570 | double average_util; | |
5571 | double min_var; | |
5572 | double max_var; | |
5573 | double stddev; | |
5574 | double sum; | |
9f95a23c | 5575 | int class_id = -1; |
11fdf7f2 TL |
5576 | set<int> allowed; |
5577 | set<int> dumped_osds; | |
31f18b77 FG |
5578 | }; |
5579 | ||
5580 | ||
5581 | class OSDUtilizationPlainDumper : public OSDUtilizationDumper<TextTable> { | |
5582 | public: | |
5583 | typedef OSDUtilizationDumper<TextTable> Parent; | |
5584 | ||
5585 | OSDUtilizationPlainDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 5586 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
5587 | const string& filter) : |
5588 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
5589 | |
5590 | void dump(TextTable *tbl) { | |
5591 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 5592 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5593 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
5594 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); | |
5595 | tbl->define_column("SIZE", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 TL |
5596 | tbl->define_column("RAW USE", TextTable::LEFT, TextTable::RIGHT); |
5597 | tbl->define_column("DATA", TextTable::LEFT, TextTable::RIGHT); | |
5598 | tbl->define_column("OMAP", TextTable::LEFT, TextTable::RIGHT); | |
5599 | tbl->define_column("META", TextTable::LEFT, TextTable::RIGHT); | |
31f18b77 FG |
5600 | tbl->define_column("AVAIL", TextTable::LEFT, TextTable::RIGHT); |
5601 | tbl->define_column("%USE", TextTable::LEFT, TextTable::RIGHT); | |
5602 | tbl->define_column("VAR", TextTable::LEFT, TextTable::RIGHT); | |
5603 | tbl->define_column("PGS", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 | 5604 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5605 | if (tree) |
5606 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
5607 | ||
5608 | Parent::dump(tbl); | |
5609 | ||
5610 | dump_stray(tbl); | |
5611 | ||
11fdf7f2 | 5612 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
224ce89b WB |
5613 | *tbl << "" |
5614 | << "" | |
5615 | << "" << "TOTAL" | |
11fdf7f2 TL |
5616 | << byte_u_t(sum.statfs.total) |
5617 | << byte_u_t(sum.statfs.get_used_raw()) | |
5618 | << byte_u_t(sum.statfs.allocated) | |
5619 | << byte_u_t(sum.statfs.omap_allocated) | |
5620 | << byte_u_t(sum.statfs.internal_metadata) | |
5621 | << byte_u_t(sum.statfs.available) | |
31f18b77 FG |
5622 | << lowprecision_t(average_util) |
5623 | << "" | |
5624 | << TextTable::endrow; | |
5625 | } | |
5626 | ||
5627 | protected: | |
5628 | struct lowprecision_t { | |
5629 | float v; | |
5630 | explicit lowprecision_t(float _v) : v(_v) {} | |
5631 | }; | |
5632 | friend std::ostream &operator<<(ostream& out, const lowprecision_t& v); | |
5633 | ||
5634 | using OSDUtilizationDumper<TextTable>::dump_item; | |
5635 | void dump_item(const CrushTreeDumper::Item &qi, | |
5636 | float &reweight, | |
5637 | int64_t kb, | |
5638 | int64_t kb_used, | |
11fdf7f2 TL |
5639 | int64_t kb_used_data, |
5640 | int64_t kb_used_omap, | |
5641 | int64_t kb_used_meta, | |
31f18b77 FG |
5642 | int64_t kb_avail, |
5643 | double& util, | |
5644 | double& var, | |
5645 | const size_t num_pgs, | |
5646 | TextTable *tbl) override { | |
224ce89b WB |
5647 | const char *c = crush->get_item_class(qi.id); |
5648 | if (!c) | |
5649 | c = ""; | |
31f18b77 | 5650 | *tbl << qi.id |
224ce89b | 5651 | << c |
31f18b77 FG |
5652 | << weightf_t(qi.weight) |
5653 | << weightf_t(reweight) | |
1adf2230 AA |
5654 | << byte_u_t(kb << 10) |
5655 | << byte_u_t(kb_used << 10) | |
11fdf7f2 TL |
5656 | << byte_u_t(kb_used_data << 10) |
5657 | << byte_u_t(kb_used_omap << 10) | |
5658 | << byte_u_t(kb_used_meta << 10) | |
1adf2230 | 5659 | << byte_u_t(kb_avail << 10) |
31f18b77 FG |
5660 | << lowprecision_t(util) |
5661 | << lowprecision_t(var); | |
5662 | ||
5663 | if (qi.is_bucket()) { | |
5664 | *tbl << "-"; | |
11fdf7f2 | 5665 | *tbl << ""; |
31f18b77 FG |
5666 | } else { |
5667 | *tbl << num_pgs; | |
11fdf7f2 TL |
5668 | if (osdmap->is_up(qi.id)) { |
5669 | *tbl << "up"; | |
5670 | } else if (osdmap->is_destroyed(qi.id)) { | |
5671 | *tbl << "destroyed"; | |
5672 | } else { | |
5673 | *tbl << "down"; | |
5674 | } | |
31f18b77 FG |
5675 | } |
5676 | ||
5677 | if (tree) { | |
5678 | ostringstream name; | |
5679 | for (int k = 0; k < qi.depth; k++) | |
5680 | name << " "; | |
5681 | if (qi.is_bucket()) { | |
5682 | int type = crush->get_bucket_type(qi.id); | |
5683 | name << crush->get_type_name(type) << " " | |
5684 | << crush->get_item_name(qi.id); | |
5685 | } else { | |
5686 | name << "osd." << qi.id; | |
5687 | } | |
5688 | *tbl << name.str(); | |
5689 | } | |
5690 | ||
5691 | *tbl << TextTable::endrow; | |
5692 | } | |
5693 | ||
5694 | public: | |
5695 | string summary() { | |
5696 | ostringstream out; | |
5697 | out << "MIN/MAX VAR: " << lowprecision_t(min_var) | |
5698 | << "/" << lowprecision_t(max_var) << " " | |
5699 | << "STDDEV: " << lowprecision_t(dev()); | |
5700 | return out.str(); | |
5701 | } | |
5702 | }; | |
5703 | ||
5704 | ostream& operator<<(ostream& out, | |
5705 | const OSDUtilizationPlainDumper::lowprecision_t& v) | |
5706 | { | |
5707 | if (v.v < -0.01) { | |
5708 | return out << "-"; | |
5709 | } else if (v.v < 0.001) { | |
5710 | return out << "0"; | |
5711 | } else { | |
5712 | std::streamsize p = out.precision(); | |
5713 | return out << std::fixed << std::setprecision(2) << v.v << std::setprecision(p); | |
5714 | } | |
5715 | } | |
5716 | ||
5717 | class OSDUtilizationFormatDumper : public OSDUtilizationDumper<Formatter> { | |
5718 | public: | |
5719 | typedef OSDUtilizationDumper<Formatter> Parent; | |
5720 | ||
5721 | OSDUtilizationFormatDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 5722 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
5723 | const string& filter) : |
5724 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
5725 | |
5726 | void dump(Formatter *f) { | |
5727 | f->open_array_section("nodes"); | |
5728 | Parent::dump(f); | |
5729 | f->close_section(); | |
5730 | ||
5731 | f->open_array_section("stray"); | |
5732 | dump_stray(f); | |
5733 | f->close_section(); | |
5734 | } | |
5735 | ||
5736 | protected: | |
5737 | using OSDUtilizationDumper<Formatter>::dump_item; | |
5738 | void dump_item(const CrushTreeDumper::Item &qi, | |
11fdf7f2 TL |
5739 | float &reweight, |
5740 | int64_t kb, | |
5741 | int64_t kb_used, | |
5742 | int64_t kb_used_data, | |
5743 | int64_t kb_used_omap, | |
5744 | int64_t kb_used_meta, | |
5745 | int64_t kb_avail, | |
5746 | double& util, | |
5747 | double& var, | |
5748 | const size_t num_pgs, | |
5749 | Formatter *f) override { | |
31f18b77 | 5750 | f->open_object_section("item"); |
c07f9fc5 | 5751 | CrushTreeDumper::dump_item_fields(crush, weight_set_names, qi, f); |
31f18b77 FG |
5752 | f->dump_float("reweight", reweight); |
5753 | f->dump_int("kb", kb); | |
5754 | f->dump_int("kb_used", kb_used); | |
11fdf7f2 TL |
5755 | f->dump_int("kb_used_data", kb_used_data); |
5756 | f->dump_int("kb_used_omap", kb_used_omap); | |
5757 | f->dump_int("kb_used_meta", kb_used_meta); | |
31f18b77 FG |
5758 | f->dump_int("kb_avail", kb_avail); |
5759 | f->dump_float("utilization", util); | |
5760 | f->dump_float("var", var); | |
5761 | f->dump_unsigned("pgs", num_pgs); | |
11fdf7f2 TL |
5762 | if (!qi.is_bucket()) { |
5763 | if (osdmap->is_up(qi.id)) { | |
5764 | f->dump_string("status", "up"); | |
5765 | } else if (osdmap->is_destroyed(qi.id)) { | |
5766 | f->dump_string("status", "destroyed"); | |
5767 | } else { | |
5768 | f->dump_string("status", "down"); | |
5769 | } | |
5770 | } | |
31f18b77 FG |
5771 | CrushTreeDumper::dump_bucket_children(crush, qi, f); |
5772 | f->close_section(); | |
5773 | } | |
5774 | ||
5775 | public: | |
5776 | void summary(Formatter *f) { | |
5777 | f->open_object_section("summary"); | |
11fdf7f2 TL |
5778 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
5779 | auto& s = sum.statfs; | |
5780 | ||
5781 | f->dump_int("total_kb", s.kb()); | |
5782 | f->dump_int("total_kb_used", s.kb_used_raw()); | |
5783 | f->dump_int("total_kb_used_data", s.kb_used_data()); | |
5784 | f->dump_int("total_kb_used_omap", s.kb_used_omap()); | |
5785 | f->dump_int("total_kb_used_meta", s.kb_used_internal_metadata()); | |
5786 | f->dump_int("total_kb_avail", s.kb_avail()); | |
31f18b77 FG |
5787 | f->dump_float("average_utilization", average_util); |
5788 | f->dump_float("min_var", min_var); | |
5789 | f->dump_float("max_var", max_var); | |
5790 | f->dump_float("dev", dev()); | |
5791 | f->close_section(); | |
5792 | } | |
5793 | }; | |
5794 | ||
5795 | void print_osd_utilization(const OSDMap& osdmap, | |
11fdf7f2 TL |
5796 | const PGMap& pgmap, |
5797 | ostream& out, | |
5798 | Formatter *f, | |
5799 | bool tree, | |
9f95a23c | 5800 | const string& filter) |
31f18b77 FG |
5801 | { |
5802 | const CrushWrapper *crush = osdmap.crush.get(); | |
5803 | if (f) { | |
5804 | f->open_object_section("df"); | |
9f95a23c | 5805 | OSDUtilizationFormatDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
5806 | d.dump(f); |
5807 | d.summary(f); | |
5808 | f->close_section(); | |
5809 | f->flush(out); | |
5810 | } else { | |
9f95a23c | 5811 | OSDUtilizationPlainDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
5812 | TextTable tbl; |
5813 | d.dump(&tbl); | |
5814 | out << tbl << d.summary() << "\n"; | |
5815 | } | |
5816 | } | |
224ce89b | 5817 | |
92f5a8d4 TL |
5818 | void OSDMap::check_health(CephContext *cct, |
5819 | health_check_map_t *checks) const | |
224ce89b WB |
5820 | { |
5821 | int num_osds = get_num_osds(); | |
5822 | ||
5823 | // OSD_DOWN | |
5824 | // OSD_$subtree_DOWN | |
5825 | // OSD_ORPHAN | |
5826 | if (num_osds >= 0) { | |
5827 | int num_in_osds = 0; | |
5828 | int num_down_in_osds = 0; | |
5829 | set<int> osds; | |
5830 | set<int> down_in_osds; | |
5831 | set<int> up_in_osds; | |
5832 | set<int> subtree_up; | |
5833 | unordered_map<int, set<int> > subtree_type_down; | |
5834 | unordered_map<int, int> num_osds_subtree; | |
5835 | int max_type = crush->get_max_type_id(); | |
5836 | ||
5837 | for (int i = 0; i < get_max_osd(); i++) { | |
5838 | if (!exists(i)) { | |
5839 | if (crush->item_exists(i)) { | |
5840 | osds.insert(i); | |
5841 | } | |
5842 | continue; | |
5843 | } | |
f67539c2 | 5844 | if (is_out(i) || (osd_state[i] & CEPH_OSD_NEW)) |
224ce89b WB |
5845 | continue; |
5846 | ++num_in_osds; | |
5847 | if (down_in_osds.count(i) || up_in_osds.count(i)) | |
5848 | continue; | |
5849 | if (!is_up(i)) { | |
5850 | down_in_osds.insert(i); | |
5851 | int parent_id = 0; | |
5852 | int current = i; | |
5853 | for (int type = 0; type <= max_type; type++) { | |
5854 | if (!crush->get_type_name(type)) | |
5855 | continue; | |
5856 | int r = crush->get_immediate_parent_id(current, &parent_id); | |
5857 | if (r == -ENOENT) | |
5858 | break; | |
5859 | // break early if this parent is already marked as up | |
5860 | if (subtree_up.count(parent_id)) | |
5861 | break; | |
5862 | type = crush->get_bucket_type(parent_id); | |
5863 | if (!subtree_type_is_down( | |
92f5a8d4 | 5864 | cct, parent_id, type, |
224ce89b WB |
5865 | &down_in_osds, &up_in_osds, &subtree_up, &subtree_type_down)) |
5866 | break; | |
5867 | current = parent_id; | |
5868 | } | |
5869 | } | |
5870 | } | |
5871 | ||
5872 | // calculate the number of down osds in each down subtree and | |
5873 | // store it in num_osds_subtree | |
5874 | for (int type = 1; type <= max_type; type++) { | |
5875 | if (!crush->get_type_name(type)) | |
5876 | continue; | |
5877 | for (auto j = subtree_type_down[type].begin(); | |
5878 | j != subtree_type_down[type].end(); | |
5879 | ++j) { | |
5880 | list<int> children; | |
5881 | int num = 0; | |
5882 | int num_children = crush->get_children(*j, &children); | |
5883 | if (num_children == 0) | |
5884 | continue; | |
5885 | for (auto l = children.begin(); l != children.end(); ++l) { | |
5886 | if (*l >= 0) { | |
5887 | ++num; | |
5888 | } else if (num_osds_subtree[*l] > 0) { | |
5889 | num = num + num_osds_subtree[*l]; | |
5890 | } | |
5891 | } | |
5892 | num_osds_subtree[*j] = num; | |
5893 | } | |
5894 | } | |
5895 | num_down_in_osds = down_in_osds.size(); | |
11fdf7f2 | 5896 | ceph_assert(num_down_in_osds <= num_in_osds); |
224ce89b WB |
5897 | if (num_down_in_osds > 0) { |
5898 | // summary of down subtree types and osds | |
5899 | for (int type = max_type; type > 0; type--) { | |
5900 | if (!crush->get_type_name(type)) | |
5901 | continue; | |
5902 | if (subtree_type_down[type].size() > 0) { | |
5903 | ostringstream ss; | |
5904 | ss << subtree_type_down[type].size() << " " | |
5905 | << crush->get_type_name(type); | |
5906 | if (subtree_type_down[type].size() > 1) { | |
5907 | ss << "s"; | |
5908 | } | |
5909 | int sum_down_osds = 0; | |
5910 | for (auto j = subtree_type_down[type].begin(); | |
5911 | j != subtree_type_down[type].end(); | |
5912 | ++j) { | |
5913 | sum_down_osds = sum_down_osds + num_osds_subtree[*j]; | |
5914 | } | |
5915 | ss << " (" << sum_down_osds << " osds) down"; | |
5916 | string err = string("OSD_") + | |
5917 | string(crush->get_type_name(type)) + "_DOWN"; | |
5918 | boost::to_upper(err); | |
9f95a23c TL |
5919 | auto& d = checks->add(err, HEALTH_WARN, ss.str(), |
5920 | subtree_type_down[type].size()); | |
224ce89b WB |
5921 | for (auto j = subtree_type_down[type].rbegin(); |
5922 | j != subtree_type_down[type].rend(); | |
5923 | ++j) { | |
5924 | ostringstream ss; | |
5925 | ss << crush->get_type_name(type); | |
5926 | ss << " "; | |
5927 | ss << crush->get_item_name(*j); | |
5928 | // at the top level, do not print location | |
5929 | if (type != max_type) { | |
5930 | ss << " ("; | |
5931 | ss << crush->get_full_location_ordered_string(*j); | |
5932 | ss << ")"; | |
5933 | } | |
5934 | int num = num_osds_subtree[*j]; | |
5935 | ss << " (" << num << " osds)"; | |
5936 | ss << " is down"; | |
5937 | d.detail.push_back(ss.str()); | |
5938 | } | |
5939 | } | |
5940 | } | |
5941 | ostringstream ss; | |
5942 | ss << down_in_osds.size() << " osds down"; | |
9f95a23c TL |
5943 | auto& d = checks->add("OSD_DOWN", HEALTH_WARN, ss.str(), |
5944 | down_in_osds.size()); | |
224ce89b WB |
5945 | for (auto it = down_in_osds.begin(); it != down_in_osds.end(); ++it) { |
5946 | ostringstream ss; | |
5947 | ss << "osd." << *it << " ("; | |
5948 | ss << crush->get_full_location_ordered_string(*it); | |
5949 | ss << ") is down"; | |
5950 | d.detail.push_back(ss.str()); | |
5951 | } | |
5952 | } | |
5953 | ||
5954 | if (!osds.empty()) { | |
5955 | ostringstream ss; | |
5956 | ss << osds.size() << " osds exist in the crush map but not in the osdmap"; | |
9f95a23c TL |
5957 | auto& d = checks->add("OSD_ORPHAN", HEALTH_WARN, ss.str(), |
5958 | osds.size()); | |
224ce89b WB |
5959 | for (auto osd : osds) { |
5960 | ostringstream ss; | |
5961 | ss << "osd." << osd << " exists in crush map but not in osdmap"; | |
5962 | d.detail.push_back(ss.str()); | |
5963 | } | |
5964 | } | |
5965 | } | |
5966 | ||
eafe8130 TL |
5967 | std::list<std::string> scrub_messages; |
5968 | bool noscrub = false, nodeepscrub = false; | |
5969 | for (const auto &p : pools) { | |
5970 | if (p.second.flags & pg_pool_t::FLAG_NOSCRUB) { | |
5971 | ostringstream ss; | |
5972 | ss << "Pool " << get_pool_name(p.first) << " has noscrub flag"; | |
5973 | scrub_messages.push_back(ss.str()); | |
5974 | noscrub = true; | |
5975 | } | |
5976 | if (p.second.flags & pg_pool_t::FLAG_NODEEP_SCRUB) { | |
5977 | ostringstream ss; | |
5978 | ss << "Pool " << get_pool_name(p.first) << " has nodeep-scrub flag"; | |
5979 | scrub_messages.push_back(ss.str()); | |
5980 | nodeepscrub = true; | |
5981 | } | |
5982 | } | |
5983 | if (noscrub || nodeepscrub) { | |
5984 | string out = ""; | |
5985 | out += noscrub ? string("noscrub") + (nodeepscrub ? ", " : "") : ""; | |
5986 | out += nodeepscrub ? "nodeep-scrub" : ""; | |
5987 | auto& d = checks->add("POOL_SCRUB_FLAGS", HEALTH_OK, | |
9f95a23c | 5988 | "Some pool(s) have the " + out + " flag(s) set", 0); |
eafe8130 TL |
5989 | d.detail.splice(d.detail.end(), scrub_messages); |
5990 | } | |
5991 | ||
224ce89b WB |
5992 | // OSD_OUT_OF_ORDER_FULL |
5993 | { | |
5994 | // An osd could configure failsafe ratio, to something different | |
5995 | // but for now assume it is the same here. | |
92f5a8d4 | 5996 | float fsr = cct->_conf->osd_failsafe_full_ratio; |
224ce89b WB |
5997 | if (fsr > 1.0) fsr /= 100; |
5998 | float fr = get_full_ratio(); | |
5999 | float br = get_backfillfull_ratio(); | |
6000 | float nr = get_nearfull_ratio(); | |
6001 | ||
6002 | list<string> detail; | |
6003 | // These checks correspond to how OSDService::check_full_status() in an OSD | |
6004 | // handles the improper setting of these values. | |
6005 | if (br < nr) { | |
6006 | ostringstream ss; | |
6007 | ss << "backfillfull_ratio (" << br | |
6008 | << ") < nearfull_ratio (" << nr << "), increased"; | |
6009 | detail.push_back(ss.str()); | |
6010 | br = nr; | |
6011 | } | |
6012 | if (fr < br) { | |
6013 | ostringstream ss; | |
6014 | ss << "full_ratio (" << fr << ") < backfillfull_ratio (" << br | |
6015 | << "), increased"; | |
6016 | detail.push_back(ss.str()); | |
6017 | fr = br; | |
6018 | } | |
6019 | if (fsr < fr) { | |
6020 | ostringstream ss; | |
6021 | ss << "osd_failsafe_full_ratio (" << fsr << ") < full_ratio (" << fr | |
6022 | << "), increased"; | |
6023 | detail.push_back(ss.str()); | |
6024 | } | |
6025 | if (!detail.empty()) { | |
6026 | auto& d = checks->add("OSD_OUT_OF_ORDER_FULL", HEALTH_ERR, | |
9f95a23c | 6027 | "full ratio(s) out of order", 0); |
224ce89b WB |
6028 | d.detail.swap(detail); |
6029 | } | |
6030 | } | |
6031 | ||
6032 | // OSD_FULL | |
6033 | // OSD_NEARFULL | |
6034 | // OSD_BACKFILLFULL | |
6035 | // OSD_FAILSAFE_FULL | |
6036 | { | |
6037 | set<int> full, backfillfull, nearfull; | |
6038 | get_full_osd_counts(&full, &backfillfull, &nearfull); | |
6039 | if (full.size()) { | |
6040 | ostringstream ss; | |
6041 | ss << full.size() << " full osd(s)"; | |
9f95a23c | 6042 | auto& d = checks->add("OSD_FULL", HEALTH_ERR, ss.str(), full.size()); |
224ce89b WB |
6043 | for (auto& i: full) { |
6044 | ostringstream ss; | |
6045 | ss << "osd." << i << " is full"; | |
6046 | d.detail.push_back(ss.str()); | |
6047 | } | |
6048 | } | |
6049 | if (backfillfull.size()) { | |
6050 | ostringstream ss; | |
6051 | ss << backfillfull.size() << " backfillfull osd(s)"; | |
9f95a23c TL |
6052 | auto& d = checks->add("OSD_BACKFILLFULL", HEALTH_WARN, ss.str(), |
6053 | backfillfull.size()); | |
224ce89b WB |
6054 | for (auto& i: backfillfull) { |
6055 | ostringstream ss; | |
6056 | ss << "osd." << i << " is backfill full"; | |
6057 | d.detail.push_back(ss.str()); | |
6058 | } | |
6059 | } | |
6060 | if (nearfull.size()) { | |
6061 | ostringstream ss; | |
6062 | ss << nearfull.size() << " nearfull osd(s)"; | |
9f95a23c | 6063 | auto& d = checks->add("OSD_NEARFULL", HEALTH_WARN, ss.str(), nearfull.size()); |
224ce89b WB |
6064 | for (auto& i: nearfull) { |
6065 | ostringstream ss; | |
6066 | ss << "osd." << i << " is near full"; | |
6067 | d.detail.push_back(ss.str()); | |
6068 | } | |
6069 | } | |
6070 | } | |
6071 | ||
6072 | // OSDMAP_FLAGS | |
6073 | { | |
6074 | // warn about flags | |
6075 | uint64_t warn_flags = | |
224ce89b WB |
6076 | CEPH_OSDMAP_PAUSERD | |
6077 | CEPH_OSDMAP_PAUSEWR | | |
6078 | CEPH_OSDMAP_PAUSEREC | | |
6079 | CEPH_OSDMAP_NOUP | | |
6080 | CEPH_OSDMAP_NODOWN | | |
6081 | CEPH_OSDMAP_NOIN | | |
6082 | CEPH_OSDMAP_NOOUT | | |
6083 | CEPH_OSDMAP_NOBACKFILL | | |
6084 | CEPH_OSDMAP_NORECOVER | | |
6085 | CEPH_OSDMAP_NOSCRUB | | |
6086 | CEPH_OSDMAP_NODEEP_SCRUB | | |
6087 | CEPH_OSDMAP_NOTIERAGENT | | |
11fdf7f2 | 6088 | CEPH_OSDMAP_NOSNAPTRIM | |
224ce89b WB |
6089 | CEPH_OSDMAP_NOREBALANCE; |
6090 | if (test_flag(warn_flags)) { | |
6091 | ostringstream ss; | |
9f95a23c TL |
6092 | string s = get_flag_string(get_flags() & warn_flags); |
6093 | ss << s << " flag(s) set"; | |
6094 | checks->add("OSDMAP_FLAGS", HEALTH_WARN, ss.str(), | |
6095 | s.size() /* kludgey but sufficient */); | |
224ce89b WB |
6096 | } |
6097 | } | |
6098 | ||
6099 | // OSD_FLAGS | |
6100 | { | |
6101 | list<string> detail; | |
6102 | const unsigned flags = | |
6103 | CEPH_OSD_NOUP | | |
6104 | CEPH_OSD_NOIN | | |
6105 | CEPH_OSD_NODOWN | | |
6106 | CEPH_OSD_NOOUT; | |
6107 | for (int i = 0; i < max_osd; ++i) { | |
6108 | if (osd_state[i] & flags) { | |
6109 | ostringstream ss; | |
6110 | set<string> states; | |
6111 | OSDMap::calc_state_set(osd_state[i] & flags, states); | |
6112 | ss << "osd." << i << " has flags " << states; | |
6113 | detail.push_back(ss.str()); | |
6114 | } | |
6115 | } | |
81eedcae TL |
6116 | for (auto& i : crush_node_flags) { |
6117 | if (i.second && crush->item_exists(i.first)) { | |
6118 | ostringstream ss; | |
6119 | set<string> states; | |
6120 | OSDMap::calc_state_set(i.second, states); | |
6121 | int t = i.first >= 0 ? 0 : crush->get_bucket_type(i.first); | |
6122 | const char *tn = crush->get_type_name(t); | |
6123 | ss << (tn ? tn : "node") << " " | |
6124 | << crush->get_item_name(i.first) << " has flags " << states; | |
6125 | detail.push_back(ss.str()); | |
6126 | } | |
6127 | } | |
6128 | for (auto& i : device_class_flags) { | |
6129 | const char* class_name = crush->get_class_name(i.first); | |
6130 | if (i.second && class_name) { | |
6131 | ostringstream ss; | |
6132 | set<string> states; | |
6133 | OSDMap::calc_state_set(i.second, states); | |
6134 | ss << "device class '" << class_name << "' has flags " << states; | |
6135 | detail.push_back(ss.str()); | |
6136 | } | |
6137 | } | |
224ce89b WB |
6138 | if (!detail.empty()) { |
6139 | ostringstream ss; | |
81eedcae | 6140 | ss << detail.size() << " OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set"; |
9f95a23c | 6141 | auto& d = checks->add("OSD_FLAGS", HEALTH_WARN, ss.str(), detail.size()); |
224ce89b WB |
6142 | d.detail.swap(detail); |
6143 | } | |
6144 | } | |
6145 | ||
6146 | // OLD_CRUSH_TUNABLES | |
92f5a8d4 | 6147 | if (cct->_conf->mon_warn_on_legacy_crush_tunables) { |
224ce89b | 6148 | string min = crush->get_min_required_version(); |
92f5a8d4 | 6149 | if (min < cct->_conf->mon_crush_min_required_version) { |
224ce89b WB |
6150 | ostringstream ss; |
6151 | ss << "crush map has legacy tunables (require " << min | |
92f5a8d4 | 6152 | << ", min is " << cct->_conf->mon_crush_min_required_version << ")"; |
9f95a23c | 6153 | auto& d = checks->add("OLD_CRUSH_TUNABLES", HEALTH_WARN, ss.str(), 0); |
f67539c2 | 6154 | d.detail.push_back("see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
6155 | } |
6156 | } | |
6157 | ||
6158 | // OLD_CRUSH_STRAW_CALC_VERSION | |
92f5a8d4 | 6159 | if (cct->_conf->mon_warn_on_crush_straw_calc_version_zero) { |
224ce89b WB |
6160 | if (crush->get_straw_calc_version() == 0) { |
6161 | ostringstream ss; | |
6162 | ss << "crush map has straw_calc_version=0"; | |
9f95a23c | 6163 | auto& d = checks->add("OLD_CRUSH_STRAW_CALC_VERSION", HEALTH_WARN, ss.str(), 0); |
224ce89b | 6164 | d.detail.push_back( |
f67539c2 | 6165 | "see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
6166 | } |
6167 | } | |
6168 | ||
6169 | // CACHE_POOL_NO_HIT_SET | |
92f5a8d4 | 6170 | if (cct->_conf->mon_warn_on_cache_pools_without_hit_sets) { |
224ce89b | 6171 | list<string> detail; |
9f95a23c | 6172 | for (auto p = pools.cbegin(); p != pools.cend(); ++p) { |
224ce89b WB |
6173 | const pg_pool_t& info = p->second; |
6174 | if (info.cache_mode_requires_hit_set() && | |
6175 | info.hit_set_params.get_type() == HitSet::TYPE_NONE) { | |
6176 | ostringstream ss; | |
6177 | ss << "pool '" << get_pool_name(p->first) | |
6178 | << "' with cache_mode " << info.get_cache_mode_name() | |
6179 | << " needs hit_set_type to be set but it is not"; | |
6180 | detail.push_back(ss.str()); | |
6181 | } | |
6182 | } | |
6183 | if (!detail.empty()) { | |
6184 | ostringstream ss; | |
6185 | ss << detail.size() << " cache pools are missing hit_sets"; | |
9f95a23c TL |
6186 | auto& d = checks->add("CACHE_POOL_NO_HIT_SET", HEALTH_WARN, ss.str(), |
6187 | detail.size()); | |
224ce89b WB |
6188 | d.detail.swap(detail); |
6189 | } | |
6190 | } | |
6191 | ||
6192 | // OSD_NO_SORTBITWISE | |
11fdf7f2 | 6193 | if (!test_flag(CEPH_OSDMAP_SORTBITWISE)) { |
224ce89b | 6194 | ostringstream ss; |
11fdf7f2 | 6195 | ss << "'sortbitwise' flag is not set"; |
9f95a23c | 6196 | checks->add("OSD_NO_SORTBITWISE", HEALTH_WARN, ss.str(), 0); |
224ce89b WB |
6197 | } |
6198 | ||
6199 | // OSD_UPGRADE_FINISHED | |
20effc67 TL |
6200 | if (auto require_release = pending_require_osd_release()) { |
6201 | ostringstream ss; | |
6202 | ss << "all OSDs are running " << *require_release << " or later but" | |
6203 | << " require_osd_release < " << *require_release; | |
6204 | auto& d = checks->add("OSD_UPGRADE_FINISHED", HEALTH_WARN, ss.str(), 0); | |
6205 | d.detail.push_back(ss.str()); | |
6206 | } | |
224ce89b | 6207 | |
3efd9988 | 6208 | // POOL_NEARFULL/BACKFILLFULL/FULL |
224ce89b | 6209 | { |
3efd9988 | 6210 | list<string> full_detail, backfillfull_detail, nearfull_detail; |
224ce89b WB |
6211 | for (auto it : get_pools()) { |
6212 | const pg_pool_t &pool = it.second; | |
3efd9988 | 6213 | const string& pool_name = get_pool_name(it.first); |
224ce89b | 6214 | if (pool.has_flag(pg_pool_t::FLAG_FULL)) { |
224ce89b | 6215 | stringstream ss; |
11fdf7f2 | 6216 | if (pool.has_flag(pg_pool_t::FLAG_FULL_QUOTA)) { |
3efd9988 FG |
6217 | // may run out of space too, |
6218 | // but we want EQUOTA taking precedence | |
11fdf7f2 | 6219 | ss << "pool '" << pool_name << "' is full (running out of quota)"; |
3efd9988 FG |
6220 | } else { |
6221 | ss << "pool '" << pool_name << "' is full (no space)"; | |
6222 | } | |
6223 | full_detail.push_back(ss.str()); | |
6224 | } else if (pool.has_flag(pg_pool_t::FLAG_BACKFILLFULL)) { | |
6225 | stringstream ss; | |
6226 | ss << "pool '" << pool_name << "' is backfillfull"; | |
6227 | backfillfull_detail.push_back(ss.str()); | |
6228 | } else if (pool.has_flag(pg_pool_t::FLAG_NEARFULL)) { | |
6229 | stringstream ss; | |
6230 | ss << "pool '" << pool_name << "' is nearfull"; | |
6231 | nearfull_detail.push_back(ss.str()); | |
224ce89b WB |
6232 | } |
6233 | } | |
3efd9988 | 6234 | if (!full_detail.empty()) { |
224ce89b | 6235 | ostringstream ss; |
3efd9988 | 6236 | ss << full_detail.size() << " pool(s) full"; |
9f95a23c | 6237 | auto& d = checks->add("POOL_FULL", HEALTH_WARN, ss.str(), full_detail.size()); |
3efd9988 FG |
6238 | d.detail.swap(full_detail); |
6239 | } | |
6240 | if (!backfillfull_detail.empty()) { | |
6241 | ostringstream ss; | |
6242 | ss << backfillfull_detail.size() << " pool(s) backfillfull"; | |
9f95a23c TL |
6243 | auto& d = checks->add("POOL_BACKFILLFULL", HEALTH_WARN, ss.str(), |
6244 | backfillfull_detail.size()); | |
3efd9988 FG |
6245 | d.detail.swap(backfillfull_detail); |
6246 | } | |
6247 | if (!nearfull_detail.empty()) { | |
6248 | ostringstream ss; | |
6249 | ss << nearfull_detail.size() << " pool(s) nearfull"; | |
9f95a23c TL |
6250 | auto& d = checks->add("POOL_NEARFULL", HEALTH_WARN, ss.str(), |
6251 | nearfull_detail.size()); | |
3efd9988 | 6252 | d.detail.swap(nearfull_detail); |
224ce89b WB |
6253 | } |
6254 | } | |
92f5a8d4 TL |
6255 | |
6256 | // POOL_PG_NUM_NOT_POWER_OF_TWO | |
6257 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_pg_num_not_power_of_two")) { | |
6258 | list<string> detail; | |
6259 | for (auto it : get_pools()) { | |
6260 | if (!isp2(it.second.get_pg_num_target())) { | |
6261 | ostringstream ss; | |
6262 | ss << "pool '" << get_pool_name(it.first) | |
6263 | << "' pg_num " << it.second.get_pg_num_target() | |
6264 | << " is not a power of two"; | |
6265 | detail.push_back(ss.str()); | |
6266 | } | |
6267 | } | |
6268 | if (!detail.empty()) { | |
6269 | ostringstream ss; | |
6270 | ss << detail.size() << " pool(s) have non-power-of-two pg_num"; | |
6271 | auto& d = checks->add("POOL_PG_NUM_NOT_POWER_OF_TWO", HEALTH_WARN, | |
9f95a23c TL |
6272 | ss.str(), detail.size()); |
6273 | d.detail.swap(detail); | |
6274 | } | |
6275 | } | |
6276 | ||
6277 | // POOL_NO_REDUNDANCY | |
6278 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_no_redundancy")) | |
6279 | { | |
6280 | list<string> detail; | |
6281 | for (auto it : get_pools()) { | |
6282 | if (it.second.get_size() == 1) { | |
6283 | ostringstream ss; | |
6284 | ss << "pool '" << get_pool_name(it.first) | |
6285 | << "' has no replicas configured"; | |
6286 | detail.push_back(ss.str()); | |
6287 | } | |
6288 | } | |
6289 | if (!detail.empty()) { | |
6290 | ostringstream ss; | |
6291 | ss << detail.size() << " pool(s) have no replicas configured"; | |
6292 | auto& d = checks->add("POOL_NO_REDUNDANCY", HEALTH_WARN, | |
6293 | ss.str(), detail.size()); | |
92f5a8d4 TL |
6294 | d.detail.swap(detail); |
6295 | } | |
6296 | } | |
f67539c2 TL |
6297 | |
6298 | // DEGRADED STRETCH MODE | |
6299 | if (cct->_conf.get_val<bool>("mon_warn_on_degraded_stretch_mode")) { | |
6300 | if (recovering_stretch_mode) { | |
6301 | stringstream ss; | |
6302 | ss << "We are recovering stretch mode buckets, only requiring " | |
6303 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
6304 | checks->add("RECOVERING_STRETCH_MODE", HEALTH_WARN, | |
6305 | ss.str(), 0); | |
6306 | } else if (degraded_stretch_mode) { | |
6307 | stringstream ss; | |
6308 | ss << "We are missing stretch mode buckets, only requiring " | |
6309 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
6310 | checks->add("DEGRADED_STRETCH_MODE", HEALTH_WARN, | |
6311 | ss.str(), 0); | |
6312 | } | |
6313 | } | |
224ce89b | 6314 | } |
35e4c445 FG |
6315 | |
6316 | int OSDMap::parse_osd_id_list(const vector<string>& ls, set<int> *out, | |
6317 | ostream *ss) const | |
6318 | { | |
6319 | out->clear(); | |
6320 | for (auto i = ls.begin(); i != ls.end(); ++i) { | |
6321 | if (i == ls.begin() && | |
6322 | (*i == "any" || *i == "all" || *i == "*")) { | |
6323 | get_all_osds(*out); | |
6324 | break; | |
6325 | } | |
9f95a23c | 6326 | long osd = TOPNSPC::common::parse_osd_id(i->c_str(), ss); |
35e4c445 FG |
6327 | if (osd < 0) { |
6328 | *ss << "invalid osd id '" << *i << "'"; | |
6329 | return -EINVAL; | |
6330 | } | |
6331 | out->insert(osd); | |
6332 | } | |
6333 | return 0; | |
6334 | } | |
11fdf7f2 TL |
6335 | |
6336 | void OSDMap::get_random_up_osds_by_subtree(int n, // whoami | |
6337 | string &subtree, | |
6338 | int limit, // how many | |
6339 | set<int> skip, | |
6340 | set<int> *want) const { | |
6341 | if (limit <= 0) | |
6342 | return; | |
6343 | int subtree_type = crush->get_type_id(subtree); | |
6344 | if (subtree_type < 1) | |
6345 | return; | |
6346 | vector<int> subtrees; | |
6347 | crush->get_subtree_of_type(subtree_type, &subtrees); | |
6348 | std::random_device rd; | |
6349 | std::default_random_engine rng{rd()}; | |
6350 | std::shuffle(subtrees.begin(), subtrees.end(), rng); | |
6351 | for (auto s : subtrees) { | |
6352 | if (limit <= 0) | |
6353 | break; | |
6354 | if (crush->subtree_contains(s, n)) | |
6355 | continue; | |
6356 | vector<int> osds; | |
6357 | crush->get_children_of_type(s, 0, &osds); | |
6358 | if (osds.empty()) | |
6359 | continue; | |
6360 | vector<int> up_osds; | |
6361 | for (auto o : osds) { | |
6362 | if (is_up(o) && !skip.count(o)) | |
6363 | up_osds.push_back(o); | |
6364 | } | |
6365 | if (up_osds.empty()) | |
6366 | continue; | |
6367 | auto it = up_osds.begin(); | |
6368 | std::advance(it, (n % up_osds.size())); | |
6369 | want->insert(*it); | |
6370 | --limit; | |
6371 | } | |
6372 | } | |
6373 | ||
6374 | float OSDMap::pool_raw_used_rate(int64_t poolid) const | |
6375 | { | |
6376 | const pg_pool_t *pool = get_pg_pool(poolid); | |
6377 | assert(pool != nullptr); | |
6378 | ||
6379 | switch (pool->get_type()) { | |
6380 | case pg_pool_t::TYPE_REPLICATED: | |
6381 | return pool->get_size(); | |
11fdf7f2 TL |
6382 | case pg_pool_t::TYPE_ERASURE: |
6383 | { | |
6384 | auto& ecp = | |
6385 | get_erasure_code_profile(pool->erasure_code_profile); | |
6386 | auto pm = ecp.find("m"); | |
6387 | auto pk = ecp.find("k"); | |
6388 | if (pm != ecp.end() && pk != ecp.end()) { | |
6389 | int k = atoi(pk->second.c_str()); | |
6390 | int m = atoi(pm->second.c_str()); | |
6391 | int mk = m + k; | |
6392 | ceph_assert(mk != 0); | |
6393 | ceph_assert(k != 0); | |
6394 | return (float)mk / k; | |
6395 | } else { | |
6396 | return 0.0; | |
6397 | } | |
6398 | } | |
6399 | break; | |
6400 | default: | |
6401 | ceph_abort_msg("unrecognized pool type"); | |
6402 | } | |
6403 | } | |
81eedcae TL |
6404 | |
6405 | unsigned OSDMap::get_osd_crush_node_flags(int osd) const | |
6406 | { | |
6407 | unsigned flags = 0; | |
6408 | if (!crush_node_flags.empty()) { | |
6409 | // the map will contain type -> name | |
6410 | std::map<std::string,std::string> ploc = crush->get_full_location(osd); | |
6411 | for (auto& i : ploc) { | |
6412 | int id = crush->get_item_id(i.second); | |
6413 | auto p = crush_node_flags.find(id); | |
6414 | if (p != crush_node_flags.end()) { | |
6415 | flags |= p->second; | |
6416 | } | |
6417 | } | |
6418 | } | |
6419 | return flags; | |
6420 | } | |
6421 | ||
6422 | unsigned OSDMap::get_crush_node_flags(int id) const | |
6423 | { | |
6424 | unsigned flags = 0; | |
6425 | auto it = crush_node_flags.find(id); | |
6426 | if (it != crush_node_flags.end()) | |
6427 | flags = it->second; | |
6428 | return flags; | |
6429 | } | |
6430 | ||
6431 | unsigned OSDMap::get_device_class_flags(int id) const | |
6432 | { | |
6433 | unsigned flags = 0; | |
6434 | auto it = device_class_flags.find(id); | |
6435 | if (it != device_class_flags.end()) | |
6436 | flags = it->second; | |
6437 | return flags; | |
6438 | } | |
20effc67 TL |
6439 | |
6440 | std::optional<std::string> OSDMap::pending_require_osd_release() const | |
6441 | { | |
6442 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_QUINCY) && | |
6443 | require_osd_release < ceph_release_t::quincy) { | |
6444 | return "quincy"; | |
6445 | } | |
6446 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_PACIFIC) && | |
6447 | require_osd_release < ceph_release_t::pacific) { | |
6448 | return "pacific"; | |
6449 | } | |
6450 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_OCTOPUS) && | |
6451 | require_osd_release < ceph_release_t::octopus) { | |
6452 | return "octopus"; | |
6453 | } | |
6454 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_NAUTILUS) && | |
6455 | require_osd_release < ceph_release_t::nautilus) { | |
6456 | return "nautilus"; | |
6457 | } | |
6458 | ||
6459 | return std::nullopt; | |
6460 | } |