]>
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 | 18 | #include <algorithm> |
1e59de90 | 19 | #include <bit> |
11fdf7f2 TL |
20 | #include <optional> |
21 | #include <random> | |
1e59de90 | 22 | #include <fmt/format.h> |
11fdf7f2 | 23 | |
224ce89b WB |
24 | #include <boost/algorithm/string.hpp> |
25 | ||
7c673cae | 26 | #include "OSDMap.h" |
7c673cae | 27 | #include "common/config.h" |
3efd9988 | 28 | #include "common/errno.h" |
7c673cae FG |
29 | #include "common/Formatter.h" |
30 | #include "common/TextTable.h" | |
31 | #include "include/ceph_features.h" | |
9f95a23c | 32 | #include "include/common_fwd.h" |
7c673cae FG |
33 | #include "include/str_map.h" |
34 | ||
35 | #include "common/code_environment.h" | |
224ce89b | 36 | #include "mon/health_check.h" |
7c673cae FG |
37 | |
38 | #include "crush/CrushTreeDumper.h" | |
39 | #include "common/Clock.h" | |
11fdf7f2 TL |
40 | #include "mon/PGMap.h" |
41 | ||
9f95a23c TL |
42 | using std::list; |
43 | using std::make_pair; | |
44 | using std::map; | |
45 | using std::multimap; | |
46 | using std::ostream; | |
47 | using std::ostringstream; | |
48 | using std::pair; | |
49 | using std::set; | |
50 | using std::string; | |
51 | using std::stringstream; | |
52 | using std::unordered_map; | |
53 | using std::vector; | |
54 | ||
55 | using ceph::decode; | |
56 | using ceph::encode; | |
57 | using ceph::Formatter; | |
58 | ||
7c673cae FG |
59 | #define dout_subsys ceph_subsys_osd |
60 | ||
61 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap, osdmap, osdmap); | |
62 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap::Incremental, osdmap_inc, osdmap); | |
63 | ||
64 | ||
65 | // ---------------------------------- | |
66 | // osd_info_t | |
67 | ||
68 | void osd_info_t::dump(Formatter *f) const | |
69 | { | |
70 | f->dump_int("last_clean_begin", last_clean_begin); | |
71 | f->dump_int("last_clean_end", last_clean_end); | |
72 | f->dump_int("up_from", up_from); | |
73 | f->dump_int("up_thru", up_thru); | |
74 | f->dump_int("down_at", down_at); | |
75 | f->dump_int("lost_at", lost_at); | |
76 | } | |
77 | ||
9f95a23c | 78 | void osd_info_t::encode(ceph::buffer::list& bl) const |
7c673cae | 79 | { |
11fdf7f2 | 80 | using ceph::encode; |
7c673cae | 81 | __u8 struct_v = 1; |
11fdf7f2 TL |
82 | encode(struct_v, bl); |
83 | encode(last_clean_begin, bl); | |
84 | encode(last_clean_end, bl); | |
85 | encode(up_from, bl); | |
86 | encode(up_thru, bl); | |
87 | encode(down_at, bl); | |
88 | encode(lost_at, bl); | |
7c673cae FG |
89 | } |
90 | ||
9f95a23c | 91 | void osd_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 92 | { |
11fdf7f2 | 93 | using ceph::decode; |
7c673cae | 94 | __u8 struct_v; |
11fdf7f2 TL |
95 | decode(struct_v, bl); |
96 | decode(last_clean_begin, bl); | |
97 | decode(last_clean_end, bl); | |
98 | decode(up_from, bl); | |
99 | decode(up_thru, bl); | |
100 | decode(down_at, bl); | |
101 | decode(lost_at, bl); | |
7c673cae FG |
102 | } |
103 | ||
104 | void osd_info_t::generate_test_instances(list<osd_info_t*>& o) | |
105 | { | |
106 | o.push_back(new osd_info_t); | |
107 | o.push_back(new osd_info_t); | |
108 | o.back()->last_clean_begin = 1; | |
109 | o.back()->last_clean_end = 2; | |
110 | o.back()->up_from = 30; | |
111 | o.back()->up_thru = 40; | |
112 | o.back()->down_at = 5; | |
113 | o.back()->lost_at = 6; | |
114 | } | |
115 | ||
116 | ostream& operator<<(ostream& out, const osd_info_t& info) | |
117 | { | |
118 | out << "up_from " << info.up_from | |
119 | << " up_thru " << info.up_thru | |
120 | << " down_at " << info.down_at | |
121 | << " last_clean_interval [" << info.last_clean_begin << "," << info.last_clean_end << ")"; | |
122 | if (info.lost_at) | |
123 | out << " lost_at " << info.lost_at; | |
124 | return out; | |
125 | } | |
126 | ||
127 | // ---------------------------------- | |
128 | // osd_xinfo_t | |
129 | ||
130 | void osd_xinfo_t::dump(Formatter *f) const | |
131 | { | |
132 | f->dump_stream("down_stamp") << down_stamp; | |
133 | f->dump_float("laggy_probability", laggy_probability); | |
134 | f->dump_int("laggy_interval", laggy_interval); | |
135 | f->dump_int("features", features); | |
136 | f->dump_unsigned("old_weight", old_weight); | |
9f95a23c TL |
137 | f->dump_stream("last_purged_snaps_scrub") << last_purged_snaps_scrub; |
138 | f->dump_int("dead_epoch", dead_epoch); | |
7c673cae FG |
139 | } |
140 | ||
9f95a23c | 141 | void osd_xinfo_t::encode(ceph::buffer::list& bl, uint64_t enc_features) const |
7c673cae | 142 | { |
9f95a23c TL |
143 | uint8_t v = 4; |
144 | if (!HAVE_FEATURE(enc_features, SERVER_OCTOPUS)) { | |
145 | v = 3; | |
146 | } | |
147 | ENCODE_START(v, 1, bl); | |
11fdf7f2 | 148 | encode(down_stamp, bl); |
f67539c2 | 149 | __u32 lp = laggy_probability * float(0xfffffffful); |
11fdf7f2 TL |
150 | encode(lp, bl); |
151 | encode(laggy_interval, bl); | |
152 | encode(features, bl); | |
153 | encode(old_weight, bl); | |
9f95a23c TL |
154 | if (v >= 4) { |
155 | encode(last_purged_snaps_scrub, bl); | |
156 | encode(dead_epoch, bl); | |
157 | } | |
7c673cae FG |
158 | ENCODE_FINISH(bl); |
159 | } | |
160 | ||
9f95a23c | 161 | void osd_xinfo_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 162 | { |
9f95a23c | 163 | DECODE_START(4, bl); |
11fdf7f2 | 164 | decode(down_stamp, bl); |
7c673cae | 165 | __u32 lp; |
11fdf7f2 | 166 | decode(lp, bl); |
7c673cae | 167 | laggy_probability = (float)lp / (float)0xffffffff; |
11fdf7f2 | 168 | decode(laggy_interval, bl); |
7c673cae | 169 | if (struct_v >= 2) |
11fdf7f2 | 170 | decode(features, bl); |
7c673cae FG |
171 | else |
172 | features = 0; | |
173 | if (struct_v >= 3) | |
11fdf7f2 | 174 | decode(old_weight, bl); |
7c673cae FG |
175 | else |
176 | old_weight = 0; | |
9f95a23c TL |
177 | if (struct_v >= 4) { |
178 | decode(last_purged_snaps_scrub, bl); | |
179 | decode(dead_epoch, bl); | |
180 | } else { | |
181 | dead_epoch = 0; | |
182 | } | |
7c673cae FG |
183 | DECODE_FINISH(bl); |
184 | } | |
185 | ||
186 | void osd_xinfo_t::generate_test_instances(list<osd_xinfo_t*>& o) | |
187 | { | |
188 | o.push_back(new osd_xinfo_t); | |
189 | o.push_back(new osd_xinfo_t); | |
190 | o.back()->down_stamp = utime_t(2, 3); | |
191 | o.back()->laggy_probability = .123; | |
192 | o.back()->laggy_interval = 123456; | |
193 | o.back()->old_weight = 0x7fff; | |
194 | } | |
195 | ||
196 | ostream& operator<<(ostream& out, const osd_xinfo_t& xi) | |
197 | { | |
198 | return out << "down_stamp " << xi.down_stamp | |
199 | << " laggy_probability " << xi.laggy_probability | |
200 | << " laggy_interval " << xi.laggy_interval | |
9f95a23c TL |
201 | << " old_weight " << xi.old_weight |
202 | << " last_purged_snaps_scrub " << xi.last_purged_snaps_scrub | |
203 | << " dead_epoch " << xi.dead_epoch; | |
7c673cae FG |
204 | } |
205 | ||
206 | // ---------------------------------- | |
207 | // OSDMap::Incremental | |
208 | ||
209 | int OSDMap::Incremental::get_net_marked_out(const OSDMap *previous) const | |
210 | { | |
211 | int n = 0; | |
212 | for (auto &weight : new_weight) { | |
213 | if (weight.second == CEPH_OSD_OUT && !previous->is_out(weight.first)) | |
214 | n++; // marked out | |
215 | else if (weight.second != CEPH_OSD_OUT && previous->is_out(weight.first)) | |
216 | n--; // marked in | |
217 | } | |
218 | return n; | |
219 | } | |
220 | ||
221 | int OSDMap::Incremental::get_net_marked_down(const OSDMap *previous) const | |
222 | { | |
223 | int n = 0; | |
224 | for (auto &state : new_state) { // | |
225 | if (state.second & CEPH_OSD_UP) { | |
226 | if (previous->is_up(state.first)) | |
227 | n++; // marked down | |
228 | else | |
229 | n--; // marked up | |
230 | } | |
231 | } | |
232 | return n; | |
233 | } | |
234 | ||
235 | int OSDMap::Incremental::identify_osd(uuid_d u) const | |
236 | { | |
237 | for (auto &uuid : new_uuid) | |
238 | if (uuid.second == u) | |
239 | return uuid.first; | |
240 | return -1; | |
241 | } | |
242 | ||
f67539c2 TL |
243 | int OSDMap::Incremental::propagate_base_properties_to_tiers(CephContext *cct, |
244 | const OSDMap& osdmap) | |
7c673cae | 245 | { |
11fdf7f2 | 246 | ceph_assert(epoch == osdmap.get_epoch() + 1); |
7c673cae FG |
247 | |
248 | for (auto &new_pool : new_pools) { | |
249 | if (!new_pool.second.tiers.empty()) { | |
250 | pg_pool_t& base = new_pool.second; | |
251 | ||
11fdf7f2 TL |
252 | auto new_rem_it = new_removed_snaps.find(new_pool.first); |
253 | ||
7c673cae FG |
254 | for (const auto &tier_pool : base.tiers) { |
255 | const auto &r = new_pools.find(tier_pool); | |
256 | pg_pool_t *tier = 0; | |
257 | if (r == new_pools.end()) { | |
258 | const pg_pool_t *orig = osdmap.get_pg_pool(tier_pool); | |
259 | if (!orig) { | |
260 | lderr(cct) << __func__ << " no pool " << tier_pool << dendl; | |
261 | return -EIO; | |
262 | } | |
263 | tier = get_new_pool(tier_pool, orig); | |
264 | } else { | |
265 | tier = &r->second; | |
266 | } | |
267 | if (tier->tier_of != new_pool.first) { | |
268 | lderr(cct) << __func__ << " " << r->first << " tier_of != " << new_pool.first << dendl; | |
269 | return -EIO; | |
270 | } | |
271 | ||
272 | ldout(cct, 10) << __func__ << " from " << new_pool.first << " to " | |
273 | << tier_pool << dendl; | |
274 | tier->snap_seq = base.snap_seq; | |
275 | tier->snap_epoch = base.snap_epoch; | |
276 | tier->snaps = base.snaps; | |
277 | tier->removed_snaps = base.removed_snaps; | |
11fdf7f2 TL |
278 | tier->flags |= base.flags & (pg_pool_t::FLAG_SELFMANAGED_SNAPS| |
279 | pg_pool_t::FLAG_POOL_SNAPS); | |
280 | ||
281 | if (new_rem_it != new_removed_snaps.end()) { | |
282 | new_removed_snaps[tier_pool] = new_rem_it->second; | |
283 | } | |
f67539c2 TL |
284 | |
285 | tier->application_metadata = base.application_metadata; | |
7c673cae FG |
286 | } |
287 | } | |
288 | } | |
289 | return 0; | |
290 | } | |
291 | ||
28e407b8 AA |
292 | // ---------------------------------- |
293 | // OSDMap | |
7c673cae FG |
294 | |
295 | bool OSDMap::subtree_is_down(int id, set<int> *down_cache) const | |
296 | { | |
297 | if (id >= 0) | |
298 | return is_down(id); | |
299 | ||
300 | if (down_cache && | |
301 | down_cache->count(id)) { | |
302 | return true; | |
303 | } | |
304 | ||
305 | list<int> children; | |
306 | crush->get_children(id, &children); | |
307 | for (const auto &child : children) { | |
308 | if (!subtree_is_down(child, down_cache)) { | |
309 | return false; | |
310 | } | |
311 | } | |
312 | if (down_cache) { | |
313 | down_cache->insert(id); | |
314 | } | |
315 | return true; | |
316 | } | |
317 | ||
318 | bool OSDMap::containing_subtree_is_down(CephContext *cct, int id, int subtree_type, set<int> *down_cache) const | |
319 | { | |
320 | // use a stack-local down_cache if we didn't get one from the | |
321 | // caller. then at least this particular call will avoid duplicated | |
322 | // work. | |
323 | set<int> local_down_cache; | |
324 | if (!down_cache) { | |
325 | down_cache = &local_down_cache; | |
326 | } | |
327 | ||
328 | int current = id; | |
329 | while (true) { | |
330 | int type; | |
331 | if (current >= 0) { | |
332 | type = 0; | |
333 | } else { | |
334 | type = crush->get_bucket_type(current); | |
335 | } | |
11fdf7f2 | 336 | ceph_assert(type >= 0); |
7c673cae FG |
337 | |
338 | if (!subtree_is_down(current, down_cache)) { | |
339 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = false" << dendl; | |
340 | return false; | |
341 | } | |
342 | ||
343 | // is this a big enough subtree to be marked as down? | |
344 | if (type >= subtree_type) { | |
345 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = true ... " << type << " >= " << subtree_type << dendl; | |
346 | return true; | |
347 | } | |
348 | ||
349 | int r = crush->get_immediate_parent_id(current, ¤t); | |
350 | if (r < 0) { | |
351 | return false; | |
352 | } | |
353 | } | |
354 | } | |
355 | ||
224ce89b WB |
356 | bool OSDMap::subtree_type_is_down( |
357 | CephContext *cct, | |
358 | int id, | |
359 | int subtree_type, | |
360 | set<int> *down_in_osds, | |
361 | set<int> *up_in_osds, | |
362 | set<int> *subtree_up, | |
363 | unordered_map<int, set<int> > *subtree_type_down) const | |
31f18b77 FG |
364 | { |
365 | if (id >= 0) { | |
366 | bool is_down_ret = is_down(id); | |
367 | if (!is_out(id)) { | |
368 | if (is_down_ret) { | |
369 | down_in_osds->insert(id); | |
370 | } else { | |
371 | up_in_osds->insert(id); | |
372 | } | |
373 | } | |
374 | return is_down_ret; | |
375 | } | |
376 | ||
377 | if (subtree_type_down && | |
378 | (*subtree_type_down)[subtree_type].count(id)) { | |
379 | return true; | |
380 | } | |
381 | ||
382 | list<int> children; | |
383 | crush->get_children(id, &children); | |
384 | for (const auto &child : children) { | |
224ce89b WB |
385 | if (!subtree_type_is_down( |
386 | cct, child, crush->get_bucket_type(child), | |
387 | down_in_osds, up_in_osds, subtree_up, subtree_type_down)) { | |
31f18b77 FG |
388 | subtree_up->insert(id); |
389 | return false; | |
390 | } | |
391 | } | |
392 | if (subtree_type_down) { | |
393 | (*subtree_type_down)[subtree_type].insert(id); | |
394 | } | |
395 | return true; | |
396 | } | |
397 | ||
9f95a23c | 398 | void OSDMap::Incremental::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 399 | { |
11fdf7f2 | 400 | using ceph::encode; |
7c673cae | 401 | __u16 v = 5; |
11fdf7f2 TL |
402 | encode(v, bl); |
403 | encode(fsid, bl); | |
404 | encode(epoch, bl); | |
405 | encode(modified, bl); | |
7c673cae | 406 | int32_t new_t = new_pool_max; |
11fdf7f2 TL |
407 | encode(new_t, bl); |
408 | encode(new_flags, bl); | |
409 | encode(fullmap, bl); | |
410 | encode(crush, bl); | |
7c673cae | 411 | |
11fdf7f2 TL |
412 | encode(new_max_osd, bl); |
413 | // for encode(new_pools, bl); | |
7c673cae | 414 | __u32 n = new_pools.size(); |
11fdf7f2 | 415 | encode(n, bl); |
7c673cae FG |
416 | for (const auto &new_pool : new_pools) { |
417 | n = new_pool.first; | |
11fdf7f2 TL |
418 | encode(n, bl); |
419 | encode(new_pool.second, bl, 0); | |
7c673cae | 420 | } |
11fdf7f2 | 421 | // for encode(new_pool_names, bl); |
7c673cae | 422 | n = new_pool_names.size(); |
11fdf7f2 | 423 | encode(n, bl); |
7c673cae FG |
424 | |
425 | for (const auto &new_pool_name : new_pool_names) { | |
426 | n = new_pool_name.first; | |
11fdf7f2 TL |
427 | encode(n, bl); |
428 | encode(new_pool_name.second, bl); | |
7c673cae | 429 | } |
11fdf7f2 | 430 | // for encode(old_pools, bl); |
7c673cae | 431 | n = old_pools.size(); |
11fdf7f2 | 432 | encode(n, bl); |
7c673cae FG |
433 | for (auto &old_pool : old_pools) { |
434 | n = old_pool; | |
11fdf7f2 | 435 | encode(n, bl); |
7c673cae | 436 | } |
11fdf7f2 | 437 | encode(new_up_client, bl, 0); |
31f18b77 FG |
438 | { |
439 | // legacy is map<int32_t,uint8_t> | |
9f95a23c | 440 | map<int32_t, uint8_t> os; |
31f18b77 | 441 | for (auto p : new_state) { |
9f95a23c TL |
442 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
443 | // that an old client could not understand. | |
444 | // skip those! | |
445 | uint8_t s = p.second; | |
446 | if (p.second != 0 && s == 0) | |
447 | continue; | |
448 | os[p.first] = s; | |
449 | } | |
450 | uint32_t n = os.size(); | |
451 | encode(n, bl); | |
452 | for (auto p : os) { | |
11fdf7f2 | 453 | encode(p.first, bl); |
9f95a23c | 454 | encode(p.second, bl); |
31f18b77 FG |
455 | } |
456 | } | |
11fdf7f2 TL |
457 | encode(new_weight, bl); |
458 | // for encode(new_pg_temp, bl); | |
7c673cae | 459 | n = new_pg_temp.size(); |
11fdf7f2 | 460 | encode(n, bl); |
7c673cae FG |
461 | |
462 | for (const auto &pg_temp : new_pg_temp) { | |
463 | old_pg_t opg = pg_temp.first.get_old_pg(); | |
11fdf7f2 TL |
464 | encode(opg, bl); |
465 | encode(pg_temp.second, bl); | |
7c673cae FG |
466 | } |
467 | } | |
468 | ||
9f95a23c | 469 | void OSDMap::Incremental::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 470 | { |
11fdf7f2 | 471 | using ceph::encode; |
7c673cae FG |
472 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
473 | encode_client_old(bl); | |
474 | return; | |
475 | } | |
476 | ||
477 | // base | |
478 | __u16 v = 6; | |
11fdf7f2 TL |
479 | encode(v, bl); |
480 | encode(fsid, bl); | |
481 | encode(epoch, bl); | |
482 | encode(modified, bl); | |
483 | encode(new_pool_max, bl); | |
484 | encode(new_flags, bl); | |
485 | encode(fullmap, bl); | |
486 | encode(crush, bl); | |
487 | ||
488 | encode(new_max_osd, bl); | |
489 | encode(new_pools, bl, features); | |
490 | encode(new_pool_names, bl); | |
491 | encode(old_pools, bl); | |
492 | encode(new_up_client, bl, features); | |
31f18b77 | 493 | { |
9f95a23c | 494 | map<int32_t, uint8_t> os; |
31f18b77 | 495 | for (auto p : new_state) { |
9f95a23c TL |
496 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
497 | // that an old client could not understand. | |
498 | // skip those! | |
499 | uint8_t s = p.second; | |
500 | if (p.second != 0 && s == 0) | |
501 | continue; | |
502 | os[p.first] = s; | |
503 | } | |
504 | uint32_t n = os.size(); | |
505 | encode(n, bl); | |
506 | for (auto p : os) { | |
11fdf7f2 | 507 | encode(p.first, bl); |
9f95a23c | 508 | encode(p.second, bl); |
31f18b77 FG |
509 | } |
510 | } | |
11fdf7f2 TL |
511 | encode(new_weight, bl); |
512 | encode(new_pg_temp, bl); | |
7c673cae FG |
513 | |
514 | // extended | |
515 | __u16 ev = 10; | |
11fdf7f2 TL |
516 | encode(ev, bl); |
517 | encode(new_hb_back_up, bl, features); | |
518 | encode(new_up_thru, bl); | |
519 | encode(new_last_clean_interval, bl); | |
520 | encode(new_lost, bl); | |
f67539c2 TL |
521 | encode(new_blocklist, bl, features); |
522 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
523 | encode(new_up_cluster, bl, features); |
524 | encode(cluster_snapshot, bl); | |
525 | encode(new_uuid, bl); | |
9f95a23c | 526 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
527 | encode(new_hb_front_up, bl, features); |
528 | } | |
529 | ||
530 | template<class T> | |
9f95a23c | 531 | static void encode_addrvec_map_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
532 | { |
533 | uint32_t n = m.size(); | |
534 | encode(n, bl); | |
535 | for (auto& i : m) { | |
536 | encode(i.first, bl); | |
537 | encode(i.second.legacy_addr(), bl, f); | |
538 | } | |
539 | } | |
540 | ||
541 | template<class T> | |
9f95a23c | 542 | static void encode_addrvec_pvec_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
543 | { |
544 | uint32_t n = m.size(); | |
545 | encode(n, bl); | |
546 | for (auto& i : m) { | |
547 | if (i) { | |
548 | encode(i->legacy_addr(), bl, f); | |
549 | } else { | |
550 | encode(entity_addr_t(), bl, f); | |
551 | } | |
552 | } | |
7c673cae FG |
553 | } |
554 | ||
11fdf7f2 TL |
555 | /* for a description of osdmap incremental versions, and when they were |
556 | * introduced, please refer to | |
557 | * doc/dev/osd_internals/osdmap_versions.txt | |
558 | */ | |
9f95a23c | 559 | void OSDMap::Incremental::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 560 | { |
11fdf7f2 | 561 | using ceph::encode; |
7c673cae FG |
562 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
563 | encode_classic(bl, features); | |
564 | return; | |
565 | } | |
566 | ||
567 | // only a select set of callers should *ever* be encoding new | |
568 | // OSDMaps. others should be passing around the canonical encoded | |
569 | // buffers from on high. select out those callers by passing in an | |
570 | // "impossible" feature bit. | |
11fdf7f2 | 571 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
572 | features &= ~CEPH_FEATURE_RESERVED; |
573 | ||
574 | size_t start_offset = bl.length(); | |
575 | size_t tail_offset; | |
11fdf7f2 | 576 | size_t crc_offset; |
9f95a23c | 577 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
578 | |
579 | // meta-encoding: how we include client-used and osd-specific data | |
580 | ENCODE_START(8, 7, bl); | |
581 | ||
582 | { | |
1e59de90 | 583 | uint8_t v = 9; |
7c673cae FG |
584 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
585 | v = 3; | |
11fdf7f2 TL |
586 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
587 | v = 5; | |
588 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
589 | v = 6; | |
df9f7d3d | 590 | } else if (!HAVE_FEATURE(features, SERVER_REEF)) { |
1e59de90 | 591 | v = 8; |
df9f7d3d | 592 | } |
7c673cae | 593 | ENCODE_START(v, 1, bl); // client-usable data |
11fdf7f2 TL |
594 | encode(fsid, bl); |
595 | encode(epoch, bl); | |
596 | encode(modified, bl); | |
597 | encode(new_pool_max, bl); | |
598 | encode(new_flags, bl); | |
599 | encode(fullmap, bl); | |
600 | encode(crush, bl); | |
601 | ||
602 | encode(new_max_osd, bl); | |
603 | encode(new_pools, bl, features); | |
604 | encode(new_pool_names, bl); | |
605 | encode(old_pools, bl); | |
606 | if (v >= 7) { | |
607 | encode(new_up_client, bl, features); | |
608 | } else { | |
609 | encode_addrvec_map_as_addr(new_up_client, bl, features); | |
610 | } | |
31f18b77 | 611 | if (v >= 5) { |
11fdf7f2 | 612 | encode(new_state, bl); |
31f18b77 | 613 | } else { |
9f95a23c | 614 | map<int32_t, uint8_t> os; |
31f18b77 | 615 | for (auto p : new_state) { |
9f95a23c TL |
616 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
617 | // that an old client could not understand. | |
618 | // skip those! | |
619 | uint8_t s = p.second; | |
620 | if (p.second != 0 && s == 0) | |
621 | continue; | |
622 | os[p.first] = s; | |
623 | } | |
624 | uint32_t n = os.size(); | |
625 | encode(n, bl); | |
626 | for (auto p : os) { | |
627 | encode(p.first, bl); | |
628 | encode(p.second, bl); | |
31f18b77 FG |
629 | } |
630 | } | |
11fdf7f2 TL |
631 | encode(new_weight, bl); |
632 | encode(new_pg_temp, bl); | |
633 | encode(new_primary_temp, bl); | |
634 | encode(new_primary_affinity, bl); | |
635 | encode(new_erasure_code_profiles, bl); | |
636 | encode(old_erasure_code_profiles, bl); | |
7c673cae | 637 | if (v >= 4) { |
11fdf7f2 TL |
638 | encode(new_pg_upmap, bl); |
639 | encode(old_pg_upmap, bl); | |
640 | encode(new_pg_upmap_items, bl); | |
641 | encode(old_pg_upmap_items, bl); | |
642 | } | |
643 | if (v >= 6) { | |
644 | encode(new_removed_snaps, bl); | |
645 | encode(new_purged_snaps, bl); | |
646 | } | |
647 | if (v >= 8) { | |
648 | encode(new_last_up_change, bl); | |
649 | encode(new_last_in_change, bl); | |
7c673cae | 650 | } |
1e59de90 TL |
651 | if (v >= 9) { |
652 | encode(new_pg_upmap_primary, bl); | |
653 | encode(old_pg_upmap_primary, bl); | |
654 | } | |
7c673cae FG |
655 | ENCODE_FINISH(bl); // client-usable data |
656 | } | |
657 | ||
658 | { | |
1e59de90 | 659 | uint8_t target_v = 9; // if bumping this, be aware of allow_crimson 12 |
7c673cae FG |
660 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
661 | target_v = 2; | |
11fdf7f2 TL |
662 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { |
663 | target_v = 6; | |
7c673cae | 664 | } |
f67539c2 | 665 | if (change_stretch_mode) { |
f67539c2 TL |
666 | target_v = std::max((uint8_t)10, target_v); |
667 | } | |
33c7a0ef TL |
668 | if (!new_range_blocklist.empty() || |
669 | !old_range_blocklist.empty()) { | |
670 | target_v = std::max((uint8_t)11, target_v); | |
671 | } | |
1e59de90 TL |
672 | if (mutate_allow_crimson != mutate_allow_crimson_t::NONE) { |
673 | target_v = std::max((uint8_t)12, target_v); | |
674 | } | |
7c673cae | 675 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
676 | if (target_v < 7) { |
677 | encode_addrvec_map_as_addr(new_hb_back_up, bl, features); | |
678 | } else { | |
679 | encode(new_hb_back_up, bl, features); | |
680 | } | |
681 | encode(new_up_thru, bl); | |
682 | encode(new_last_clean_interval, bl); | |
683 | encode(new_lost, bl); | |
f67539c2 TL |
684 | encode(new_blocklist, bl, features); |
685 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
686 | if (target_v < 7) { |
687 | encode_addrvec_map_as_addr(new_up_cluster, bl, features); | |
688 | } else { | |
689 | encode(new_up_cluster, bl, features); | |
690 | } | |
691 | encode(cluster_snapshot, bl); | |
692 | encode(new_uuid, bl); | |
9f95a23c | 693 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
694 | if (target_v < 7) { |
695 | encode_addrvec_map_as_addr(new_hb_front_up, bl, features); | |
696 | } else { | |
697 | encode(new_hb_front_up, bl, features); | |
698 | } | |
699 | encode(features, bl); // NOTE: features arg, not the member | |
7c673cae | 700 | if (target_v >= 3) { |
11fdf7f2 TL |
701 | encode(new_nearfull_ratio, bl); |
702 | encode(new_full_ratio, bl); | |
703 | encode(new_backfillfull_ratio, bl); | |
31f18b77 FG |
704 | } |
705 | // 5 was string-based new_require_min_compat_client | |
706 | if (target_v >= 6) { | |
11fdf7f2 TL |
707 | encode(new_require_min_compat_client, bl); |
708 | encode(new_require_osd_release, bl); | |
7c673cae | 709 | } |
81eedcae TL |
710 | if (target_v >= 8) { |
711 | encode(new_crush_node_flags, bl); | |
712 | } | |
713 | if (target_v >= 9) { | |
714 | encode(new_device_class_flags, bl); | |
715 | } | |
f67539c2 TL |
716 | if (target_v >= 10) { |
717 | encode(change_stretch_mode, bl); | |
718 | encode(new_stretch_bucket_count, bl); | |
719 | encode(new_degraded_stretch_mode, bl); | |
720 | encode(new_recovering_stretch_mode, bl); | |
721 | encode(new_stretch_mode_bucket, bl); | |
722 | encode(stretch_mode_enabled, bl); | |
723 | } | |
33c7a0ef TL |
724 | if (target_v >= 11) { |
725 | encode(new_range_blocklist, bl, features); | |
726 | encode(old_range_blocklist, bl, features); | |
727 | } | |
1e59de90 TL |
728 | if (target_v >= 12) { |
729 | encode(mutate_allow_crimson, bl); | |
730 | } | |
7c673cae FG |
731 | ENCODE_FINISH(bl); // osd-only data |
732 | } | |
733 | ||
11fdf7f2 TL |
734 | crc_offset = bl.length(); |
735 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
736 | tail_offset = bl.length(); |
737 | ||
11fdf7f2 | 738 | encode(full_crc, bl); |
7c673cae FG |
739 | |
740 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
741 | ||
742 | // fill in crc | |
9f95a23c | 743 | ceph::buffer::list front; |
11fdf7f2 | 744 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae | 745 | inc_crc = front.crc32c(-1); |
9f95a23c | 746 | ceph::buffer::list tail; |
7c673cae FG |
747 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
748 | inc_crc = tail.crc32c(inc_crc); | |
749 | ceph_le32 crc_le; | |
750 | crc_le = inc_crc; | |
11fdf7f2 | 751 | crc_filler->copy_in(4u, (char*)&crc_le); |
7c673cae FG |
752 | have_crc = true; |
753 | } | |
754 | ||
9f95a23c | 755 | void OSDMap::Incremental::decode_classic(ceph::buffer::list::const_iterator &p) |
7c673cae | 756 | { |
11fdf7f2 | 757 | using ceph::decode; |
7c673cae FG |
758 | __u32 n, t; |
759 | // base | |
760 | __u16 v; | |
11fdf7f2 TL |
761 | decode(v, p); |
762 | decode(fsid, p); | |
763 | decode(epoch, p); | |
764 | decode(modified, p); | |
7c673cae | 765 | if (v == 4 || v == 5) { |
11fdf7f2 | 766 | decode(n, p); |
7c673cae FG |
767 | new_pool_max = n; |
768 | } else if (v >= 6) | |
11fdf7f2 TL |
769 | decode(new_pool_max, p); |
770 | decode(new_flags, p); | |
771 | decode(fullmap, p); | |
772 | decode(crush, p); | |
7c673cae | 773 | |
11fdf7f2 | 774 | decode(new_max_osd, p); |
7c673cae FG |
775 | if (v < 6) { |
776 | new_pools.clear(); | |
11fdf7f2 | 777 | decode(n, p); |
7c673cae | 778 | while (n--) { |
11fdf7f2 TL |
779 | decode(t, p); |
780 | decode(new_pools[t], p); | |
7c673cae FG |
781 | } |
782 | } else { | |
11fdf7f2 | 783 | decode(new_pools, p); |
7c673cae FG |
784 | } |
785 | if (v == 5) { | |
786 | new_pool_names.clear(); | |
11fdf7f2 | 787 | decode(n, p); |
7c673cae | 788 | while (n--) { |
11fdf7f2 TL |
789 | decode(t, p); |
790 | decode(new_pool_names[t], p); | |
7c673cae FG |
791 | } |
792 | } else if (v >= 6) { | |
11fdf7f2 | 793 | decode(new_pool_names, p); |
7c673cae FG |
794 | } |
795 | if (v < 6) { | |
796 | old_pools.clear(); | |
11fdf7f2 | 797 | decode(n, p); |
7c673cae | 798 | while (n--) { |
11fdf7f2 | 799 | decode(t, p); |
7c673cae FG |
800 | old_pools.insert(t); |
801 | } | |
802 | } else { | |
11fdf7f2 | 803 | decode(old_pools, p); |
7c673cae | 804 | } |
11fdf7f2 | 805 | decode(new_up_client, p); |
31f18b77 FG |
806 | { |
807 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 808 | decode(ns, p); |
31f18b77 FG |
809 | for (auto q : ns) { |
810 | new_state[q.first] = q.second; | |
811 | } | |
812 | } | |
11fdf7f2 | 813 | decode(new_weight, p); |
7c673cae FG |
814 | |
815 | if (v < 6) { | |
816 | new_pg_temp.clear(); | |
11fdf7f2 | 817 | decode(n, p); |
7c673cae FG |
818 | while (n--) { |
819 | old_pg_t opg; | |
9f95a23c | 820 | ceph::decode_raw(opg, p); |
11fdf7f2 | 821 | decode(new_pg_temp[pg_t(opg)], p); |
7c673cae FG |
822 | } |
823 | } else { | |
11fdf7f2 | 824 | decode(new_pg_temp, p); |
7c673cae FG |
825 | } |
826 | ||
827 | // decode short map, too. | |
828 | if (v == 5 && p.end()) | |
829 | return; | |
830 | ||
831 | // extended | |
832 | __u16 ev = 0; | |
833 | if (v >= 5) | |
11fdf7f2 TL |
834 | decode(ev, p); |
835 | decode(new_hb_back_up, p); | |
7c673cae | 836 | if (v < 5) |
11fdf7f2 TL |
837 | decode(new_pool_names, p); |
838 | decode(new_up_thru, p); | |
839 | decode(new_last_clean_interval, p); | |
840 | decode(new_lost, p); | |
f67539c2 TL |
841 | decode(new_blocklist, p); |
842 | decode(old_blocklist, p); | |
7c673cae | 843 | if (ev >= 6) |
11fdf7f2 | 844 | decode(new_up_cluster, p); |
7c673cae | 845 | if (ev >= 7) |
11fdf7f2 | 846 | decode(cluster_snapshot, p); |
7c673cae | 847 | if (ev >= 8) |
11fdf7f2 | 848 | decode(new_uuid, p); |
7c673cae | 849 | if (ev >= 9) |
11fdf7f2 | 850 | decode(new_xinfo, p); |
7c673cae | 851 | if (ev >= 10) |
11fdf7f2 | 852 | decode(new_hb_front_up, p); |
7c673cae FG |
853 | } |
854 | ||
11fdf7f2 TL |
855 | /* for a description of osdmap incremental versions, and when they were |
856 | * introduced, please refer to | |
857 | * doc/dev/osd_internals/osdmap_versions.txt | |
858 | */ | |
9f95a23c | 859 | void OSDMap::Incremental::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 860 | { |
11fdf7f2 | 861 | using ceph::decode; |
7c673cae FG |
862 | /** |
863 | * Older encodings of the Incremental had a single struct_v which | |
864 | * covered the whole encoding, and was prior to our modern | |
865 | * stuff which includes a compatv and a size. So if we see | |
866 | * a struct_v < 7, we must rewind to the beginning and use our | |
867 | * classic decoder. | |
868 | */ | |
869 | size_t start_offset = bl.get_off(); | |
870 | size_t tail_offset = 0; | |
9f95a23c | 871 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
872 | |
873 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
874 | if (struct_v < 7) { | |
11fdf7f2 | 875 | bl.seek(start_offset); |
7c673cae FG |
876 | decode_classic(bl); |
877 | encode_features = 0; | |
878 | if (struct_v >= 6) | |
879 | encode_features = CEPH_FEATURE_PGID64; | |
880 | else | |
881 | encode_features = 0; | |
882 | return; | |
883 | } | |
884 | { | |
11fdf7f2 TL |
885 | DECODE_START(8, bl); // client-usable data |
886 | decode(fsid, bl); | |
887 | decode(epoch, bl); | |
888 | decode(modified, bl); | |
889 | decode(new_pool_max, bl); | |
890 | decode(new_flags, bl); | |
891 | decode(fullmap, bl); | |
892 | decode(crush, bl); | |
893 | ||
894 | decode(new_max_osd, bl); | |
895 | decode(new_pools, bl); | |
896 | decode(new_pool_names, bl); | |
897 | decode(old_pools, bl); | |
898 | decode(new_up_client, bl); | |
31f18b77 | 899 | if (struct_v >= 5) { |
11fdf7f2 | 900 | decode(new_state, bl); |
31f18b77 FG |
901 | } else { |
902 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 903 | decode(ns, bl); |
31f18b77 FG |
904 | for (auto q : ns) { |
905 | new_state[q.first] = q.second; | |
906 | } | |
907 | } | |
11fdf7f2 TL |
908 | decode(new_weight, bl); |
909 | decode(new_pg_temp, bl); | |
910 | decode(new_primary_temp, bl); | |
7c673cae | 911 | if (struct_v >= 2) |
11fdf7f2 | 912 | decode(new_primary_affinity, bl); |
7c673cae FG |
913 | else |
914 | new_primary_affinity.clear(); | |
915 | if (struct_v >= 3) { | |
11fdf7f2 TL |
916 | decode(new_erasure_code_profiles, bl); |
917 | decode(old_erasure_code_profiles, bl); | |
7c673cae FG |
918 | } else { |
919 | new_erasure_code_profiles.clear(); | |
920 | old_erasure_code_profiles.clear(); | |
921 | } | |
922 | if (struct_v >= 4) { | |
11fdf7f2 TL |
923 | decode(new_pg_upmap, bl); |
924 | decode(old_pg_upmap, bl); | |
925 | decode(new_pg_upmap_items, bl); | |
926 | decode(old_pg_upmap_items, bl); | |
927 | } | |
928 | if (struct_v >= 6) { | |
929 | decode(new_removed_snaps, bl); | |
930 | decode(new_purged_snaps, bl); | |
931 | } | |
932 | if (struct_v >= 8) { | |
933 | decode(new_last_up_change, bl); | |
934 | decode(new_last_in_change, bl); | |
7c673cae FG |
935 | } |
936 | DECODE_FINISH(bl); // client-usable data | |
937 | } | |
938 | ||
939 | { | |
f67539c2 | 940 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
941 | decode(new_hb_back_up, bl); |
942 | decode(new_up_thru, bl); | |
943 | decode(new_last_clean_interval, bl); | |
944 | decode(new_lost, bl); | |
f67539c2 TL |
945 | decode(new_blocklist, bl); |
946 | decode(old_blocklist, bl); | |
11fdf7f2 TL |
947 | decode(new_up_cluster, bl); |
948 | decode(cluster_snapshot, bl); | |
949 | decode(new_uuid, bl); | |
950 | decode(new_xinfo, bl); | |
951 | decode(new_hb_front_up, bl); | |
7c673cae | 952 | if (struct_v >= 2) |
11fdf7f2 | 953 | decode(encode_features, bl); |
7c673cae FG |
954 | else |
955 | encode_features = CEPH_FEATURE_PGID64 | CEPH_FEATURE_OSDMAP_ENC; | |
956 | if (struct_v >= 3) { | |
11fdf7f2 TL |
957 | decode(new_nearfull_ratio, bl); |
958 | decode(new_full_ratio, bl); | |
7c673cae FG |
959 | } else { |
960 | new_nearfull_ratio = -1; | |
961 | new_full_ratio = -1; | |
962 | } | |
963 | if (struct_v >= 4) { | |
11fdf7f2 | 964 | decode(new_backfillfull_ratio, bl); |
7c673cae FG |
965 | } else { |
966 | new_backfillfull_ratio = -1; | |
967 | } | |
31f18b77 FG |
968 | if (struct_v == 5) { |
969 | string r; | |
11fdf7f2 | 970 | decode(r, bl); |
31f18b77 | 971 | if (r.length()) { |
9f95a23c | 972 | new_require_min_compat_client = ceph_release_from_name(r); |
31f18b77 FG |
973 | } |
974 | } | |
975 | if (struct_v >= 6) { | |
11fdf7f2 TL |
976 | decode(new_require_min_compat_client, bl); |
977 | decode(new_require_osd_release, bl); | |
31f18b77 FG |
978 | } else { |
979 | if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_LUMINOUS)) { | |
980 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 981 | new_require_osd_release = ceph_release_t::luminous; |
31f18b77 FG |
982 | new_flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
983 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_KRAKEN)) { | |
9f95a23c | 984 | new_require_osd_release = ceph_release_t::kraken; |
31f18b77 | 985 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_JEWEL)) { |
9f95a23c | 986 | new_require_osd_release = ceph_release_t::jewel; |
31f18b77 | 987 | } else { |
9f95a23c | 988 | new_require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
989 | } |
990 | } | |
81eedcae TL |
991 | if (struct_v >= 8) { |
992 | decode(new_crush_node_flags, bl); | |
993 | } | |
994 | if (struct_v >= 9) { | |
995 | decode(new_device_class_flags, bl); | |
996 | } | |
f67539c2 TL |
997 | if (struct_v >= 10) { |
998 | decode(change_stretch_mode, bl); | |
999 | decode(new_stretch_bucket_count, bl); | |
1000 | decode(new_degraded_stretch_mode, bl); | |
1001 | decode(new_recovering_stretch_mode, bl); | |
1002 | decode(new_stretch_mode_bucket, bl); | |
1003 | decode(stretch_mode_enabled, bl); | |
1004 | } | |
33c7a0ef TL |
1005 | if (struct_v >= 11) { |
1006 | decode(new_range_blocklist, bl); | |
1007 | decode(old_range_blocklist, bl); | |
1008 | } | |
1e59de90 TL |
1009 | if (struct_v >= 12) { |
1010 | decode(mutate_allow_crimson, bl); | |
1011 | } | |
7c673cae FG |
1012 | DECODE_FINISH(bl); // osd-only data |
1013 | } | |
1014 | ||
1015 | if (struct_v >= 8) { | |
1016 | have_crc = true; | |
1017 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 1018 | decode(inc_crc, bl); |
7c673cae | 1019 | tail_offset = bl.get_off(); |
11fdf7f2 | 1020 | decode(full_crc, bl); |
7c673cae FG |
1021 | } else { |
1022 | have_crc = false; | |
1023 | full_crc = 0; | |
1024 | inc_crc = 0; | |
1025 | } | |
1026 | ||
1027 | DECODE_FINISH(bl); // wrapper | |
1028 | ||
1029 | if (have_crc) { | |
1030 | // verify crc | |
1031 | uint32_t actual = crc_front.crc32c(-1); | |
1032 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 1033 | ceph::buffer::list tail; |
7c673cae FG |
1034 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
1035 | actual = tail.crc32c(actual); | |
1036 | } | |
1037 | if (inc_crc != actual) { | |
1038 | ostringstream ss; | |
1039 | ss << "bad crc, actual " << actual << " != expected " << inc_crc; | |
1040 | string s = ss.str(); | |
9f95a23c | 1041 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
1042 | } |
1043 | } | |
1044 | } | |
1045 | ||
1046 | void OSDMap::Incremental::dump(Formatter *f) const | |
1047 | { | |
1048 | f->dump_int("epoch", epoch); | |
1049 | f->dump_stream("fsid") << fsid; | |
1050 | f->dump_stream("modified") << modified; | |
11fdf7f2 TL |
1051 | f->dump_stream("new_last_up_change") << new_last_up_change; |
1052 | f->dump_stream("new_last_in_change") << new_last_in_change; | |
7c673cae FG |
1053 | f->dump_int("new_pool_max", new_pool_max); |
1054 | f->dump_int("new_flags", new_flags); | |
1055 | f->dump_float("new_full_ratio", new_full_ratio); | |
1056 | f->dump_float("new_nearfull_ratio", new_nearfull_ratio); | |
1057 | f->dump_float("new_backfillfull_ratio", new_backfillfull_ratio); | |
f67539c2 TL |
1058 | f->dump_int("new_require_min_compat_client", to_integer<int>(new_require_min_compat_client)); |
1059 | f->dump_int("new_require_osd_release", to_integer<int>(new_require_osd_release)); | |
1e59de90 | 1060 | f->dump_unsigned("mutate_allow_crimson", static_cast<unsigned>(mutate_allow_crimson)); |
7c673cae FG |
1061 | |
1062 | if (fullmap.length()) { | |
1063 | f->open_object_section("full_map"); | |
1064 | OSDMap full; | |
9f95a23c | 1065 | ceph::buffer::list fbl = fullmap; // kludge around constness. |
11fdf7f2 | 1066 | auto p = fbl.cbegin(); |
7c673cae FG |
1067 | full.decode(p); |
1068 | full.dump(f); | |
1069 | f->close_section(); | |
1070 | } | |
1071 | if (crush.length()) { | |
1072 | f->open_object_section("crush"); | |
1073 | CrushWrapper c; | |
9f95a23c | 1074 | ceph::buffer::list tbl = crush; // kludge around constness. |
11fdf7f2 | 1075 | auto p = tbl.cbegin(); |
7c673cae FG |
1076 | c.decode(p); |
1077 | c.dump(f); | |
1078 | f->close_section(); | |
1079 | } | |
1080 | ||
1081 | f->dump_int("new_max_osd", new_max_osd); | |
1082 | ||
1083 | f->open_array_section("new_pools"); | |
1084 | ||
1085 | for (const auto &new_pool : new_pools) { | |
1086 | f->open_object_section("pool"); | |
1087 | f->dump_int("pool", new_pool.first); | |
1088 | new_pool.second.dump(f); | |
1089 | f->close_section(); | |
1090 | } | |
1091 | f->close_section(); | |
1092 | f->open_array_section("new_pool_names"); | |
1093 | ||
1094 | for (const auto &new_pool_name : new_pool_names) { | |
1095 | f->open_object_section("pool_name"); | |
1096 | f->dump_int("pool", new_pool_name.first); | |
1097 | f->dump_string("name", new_pool_name.second); | |
1098 | f->close_section(); | |
1099 | } | |
1100 | f->close_section(); | |
1101 | f->open_array_section("old_pools"); | |
1102 | ||
1103 | for (const auto &old_pool : old_pools) | |
1104 | f->dump_int("pool", old_pool); | |
1105 | f->close_section(); | |
1106 | ||
1107 | f->open_array_section("new_up_osds"); | |
1108 | ||
1109 | for (const auto &upclient : new_up_client) { | |
1110 | f->open_object_section("osd"); | |
1111 | f->dump_int("osd", upclient.first); | |
11fdf7f2 TL |
1112 | f->dump_stream("public_addr") << upclient.second.legacy_addr(); |
1113 | f->dump_object("public_addrs", upclient.second); | |
1114 | if (auto p = new_up_cluster.find(upclient.first); | |
1115 | p != new_up_cluster.end()) { | |
1116 | f->dump_stream("cluster_addr") << p->second.legacy_addr(); | |
1117 | f->dump_object("cluster_addrs", p->second); | |
1118 | } | |
1119 | if (auto p = new_hb_back_up.find(upclient.first); | |
1120 | p != new_hb_back_up.end()) { | |
1121 | f->dump_object("heartbeat_back_addrs", p->second); | |
1122 | } | |
1123 | if (auto p = new_hb_front_up.find(upclient.first); | |
1124 | p != new_hb_front_up.end()) { | |
1125 | f->dump_object("heartbeat_front_addrs", p->second); | |
1126 | } | |
7c673cae FG |
1127 | f->close_section(); |
1128 | } | |
1129 | f->close_section(); | |
1130 | ||
1131 | f->open_array_section("new_weight"); | |
1132 | ||
1133 | for (const auto &weight : new_weight) { | |
1134 | f->open_object_section("osd"); | |
1135 | f->dump_int("osd", weight.first); | |
1136 | f->dump_int("weight", weight.second); | |
1137 | f->close_section(); | |
1138 | } | |
1139 | f->close_section(); | |
1140 | ||
1141 | f->open_array_section("osd_state_xor"); | |
1142 | for (const auto &ns : new_state) { | |
1143 | f->open_object_section("osd"); | |
1144 | f->dump_int("osd", ns.first); | |
1145 | set<string> st; | |
1146 | calc_state_set(new_state.find(ns.first)->second, st); | |
1147 | f->open_array_section("state_xor"); | |
1148 | for (auto &state : st) | |
1149 | f->dump_string("state", state); | |
1150 | f->close_section(); | |
c07f9fc5 | 1151 | f->close_section(); |
7c673cae FG |
1152 | } |
1153 | f->close_section(); | |
1154 | ||
1155 | f->open_array_section("new_pg_temp"); | |
1156 | ||
1157 | for (const auto &pg_temp : new_pg_temp) { | |
1158 | f->open_object_section("pg"); | |
1159 | f->dump_stream("pgid") << pg_temp.first; | |
1160 | f->open_array_section("osds"); | |
1161 | ||
1162 | for (const auto &osd : pg_temp.second) | |
1163 | f->dump_int("osd", osd); | |
1164 | f->close_section(); | |
1165 | f->close_section(); | |
1166 | } | |
1167 | f->close_section(); | |
1168 | ||
1169 | f->open_array_section("primary_temp"); | |
1170 | ||
1171 | for (const auto &primary_temp : new_primary_temp) { | |
1172 | f->dump_stream("pgid") << primary_temp.first; | |
1173 | f->dump_int("osd", primary_temp.second); | |
1174 | } | |
1175 | f->close_section(); // primary_temp | |
1176 | ||
1177 | f->open_array_section("new_pg_upmap"); | |
1178 | for (auto& i : new_pg_upmap) { | |
1179 | f->open_object_section("mapping"); | |
1180 | f->dump_stream("pgid") << i.first; | |
1181 | f->open_array_section("osds"); | |
1182 | for (auto osd : i.second) { | |
1183 | f->dump_int("osd", osd); | |
1184 | } | |
1185 | f->close_section(); | |
1186 | f->close_section(); | |
1187 | } | |
1188 | f->close_section(); | |
1189 | f->open_array_section("old_pg_upmap"); | |
1190 | for (auto& i : old_pg_upmap) { | |
1191 | f->dump_stream("pgid") << i; | |
1192 | } | |
1193 | f->close_section(); | |
1194 | ||
1195 | f->open_array_section("new_pg_upmap_items"); | |
1196 | for (auto& i : new_pg_upmap_items) { | |
1197 | f->open_object_section("mapping"); | |
1198 | f->dump_stream("pgid") << i.first; | |
1199 | f->open_array_section("mappings"); | |
1200 | for (auto& p : i.second) { | |
1201 | f->open_object_section("mapping"); | |
1202 | f->dump_int("from", p.first); | |
1203 | f->dump_int("to", p.second); | |
1204 | f->close_section(); | |
1205 | } | |
1206 | f->close_section(); | |
1207 | f->close_section(); | |
1208 | } | |
1209 | f->close_section(); | |
1210 | f->open_array_section("old_pg_upmap_items"); | |
1211 | for (auto& i : old_pg_upmap_items) { | |
1212 | f->dump_stream("pgid") << i; | |
1213 | } | |
1214 | f->close_section(); | |
1215 | ||
1e59de90 TL |
1216 | // dump upmap_primaries |
1217 | f->open_array_section("new_pg_upmap_primaries"); | |
1218 | for (auto& [pg, osd] : new_pg_upmap_primary) { | |
1219 | f->open_object_section("primary_mapping"); | |
1220 | f->dump_stream("pgid") << pg; | |
1221 | f->dump_int("primary_osd", osd); | |
1222 | f->close_section(); | |
1223 | } | |
1224 | f->close_section(); // new_pg_upmap_primaries | |
1225 | ||
1226 | // dump old_pg_upmap_primaries (removed primary mappings) | |
1227 | f->open_array_section("old_pg_upmap_primaries"); | |
1228 | for (auto& pg : old_pg_upmap_primary) { | |
1229 | f->dump_stream("pgid") << pg; | |
1230 | } | |
1231 | f->close_section(); // old_pg_upmap_primaries | |
1232 | ||
7c673cae FG |
1233 | f->open_array_section("new_up_thru"); |
1234 | ||
1235 | for (const auto &up_thru : new_up_thru) { | |
1236 | f->open_object_section("osd"); | |
1237 | f->dump_int("osd", up_thru.first); | |
1238 | f->dump_int("up_thru", up_thru.second); | |
1239 | f->close_section(); | |
1240 | } | |
1241 | f->close_section(); | |
1242 | ||
1243 | f->open_array_section("new_lost"); | |
1244 | ||
1245 | for (const auto &lost : new_lost) { | |
1246 | f->open_object_section("osd"); | |
1247 | f->dump_int("osd", lost.first); | |
1248 | f->dump_int("epoch_lost", lost.second); | |
1249 | f->close_section(); | |
1250 | } | |
1251 | f->close_section(); | |
1252 | ||
1253 | f->open_array_section("new_last_clean_interval"); | |
1254 | ||
1255 | for (const auto &last_clean_interval : new_last_clean_interval) { | |
1256 | f->open_object_section("osd"); | |
1257 | f->dump_int("osd", last_clean_interval.first); | |
1258 | f->dump_int("first", last_clean_interval.second.first); | |
1259 | f->dump_int("last", last_clean_interval.second.second); | |
1260 | f->close_section(); | |
1261 | } | |
1262 | f->close_section(); | |
1263 | ||
f67539c2 TL |
1264 | f->open_array_section("new_blocklist"); |
1265 | for (const auto &blist : new_blocklist) { | |
7c673cae FG |
1266 | stringstream ss; |
1267 | ss << blist.first; | |
1268 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1269 | } | |
1270 | f->close_section(); | |
f67539c2 TL |
1271 | f->open_array_section("old_blocklist"); |
1272 | for (const auto &blist : old_blocklist) | |
7c673cae FG |
1273 | f->dump_stream("addr") << blist; |
1274 | f->close_section(); | |
33c7a0ef TL |
1275 | f->open_array_section("new_range_blocklist"); |
1276 | for (const auto &blist : new_range_blocklist) { | |
1277 | stringstream ss; | |
1278 | ss << blist.first; | |
1279 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1280 | } | |
1281 | f->close_section(); | |
1282 | f->open_array_section("old_range_blocklist"); | |
1283 | for (const auto &blist : old_range_blocklist) | |
1284 | f->dump_stream("addr") << blist; | |
1285 | f->close_section(); | |
7c673cae FG |
1286 | |
1287 | f->open_array_section("new_xinfo"); | |
1288 | for (const auto &xinfo : new_xinfo) { | |
1289 | f->open_object_section("xinfo"); | |
1290 | f->dump_int("osd", xinfo.first); | |
1291 | xinfo.second.dump(f); | |
1292 | f->close_section(); | |
1293 | } | |
1294 | f->close_section(); | |
1295 | ||
1296 | if (cluster_snapshot.size()) | |
1297 | f->dump_string("cluster_snapshot", cluster_snapshot); | |
1298 | ||
1299 | f->open_array_section("new_uuid"); | |
1300 | for (const auto &uuid : new_uuid) { | |
1301 | f->open_object_section("osd"); | |
1302 | f->dump_int("osd", uuid.first); | |
1303 | f->dump_stream("uuid") << uuid.second; | |
1304 | f->close_section(); | |
1305 | } | |
1306 | f->close_section(); | |
1307 | ||
1308 | OSDMap::dump_erasure_code_profiles(new_erasure_code_profiles, f); | |
1309 | f->open_array_section("old_erasure_code_profiles"); | |
1310 | for (const auto &erasure_code_profile : old_erasure_code_profiles) { | |
9f95a23c | 1311 | f->dump_string("old", erasure_code_profile); |
7c673cae FG |
1312 | } |
1313 | f->close_section(); | |
11fdf7f2 TL |
1314 | |
1315 | f->open_array_section("new_removed_snaps"); | |
1316 | for (auto& p : new_removed_snaps) { | |
1317 | f->open_object_section("pool"); | |
1318 | f->dump_int("pool", p.first); | |
1319 | f->open_array_section("snaps"); | |
1320 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1321 | f->open_object_section("interval"); | |
1322 | f->dump_unsigned("begin", q.get_start()); | |
1323 | f->dump_unsigned("length", q.get_len()); | |
1324 | f->close_section(); | |
1325 | } | |
1326 | f->close_section(); | |
1327 | f->close_section(); | |
1328 | } | |
1329 | f->close_section(); | |
1330 | f->open_array_section("new_purged_snaps"); | |
1331 | for (auto& p : new_purged_snaps) { | |
1332 | f->open_object_section("pool"); | |
1333 | f->dump_int("pool", p.first); | |
1334 | f->open_array_section("snaps"); | |
1335 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1336 | f->open_object_section("interval"); | |
1337 | f->dump_unsigned("begin", q.get_start()); | |
1338 | f->dump_unsigned("length", q.get_len()); | |
1339 | f->close_section(); | |
1340 | } | |
1341 | f->close_section(); | |
1342 | f->close_section(); | |
1343 | } | |
81eedcae TL |
1344 | f->open_array_section("new_crush_node_flags"); |
1345 | for (auto& i : new_crush_node_flags) { | |
1346 | f->open_object_section("node"); | |
1347 | f->dump_int("id", i.first); | |
1348 | set<string> st; | |
1349 | calc_state_set(i.second, st); | |
1350 | for (auto& j : st) { | |
1351 | f->dump_string("flag", j); | |
1352 | } | |
1353 | f->close_section(); | |
1354 | } | |
1355 | f->close_section(); | |
1356 | f->open_array_section("new_device_class_flags"); | |
1357 | for (auto& i : new_device_class_flags) { | |
1358 | f->open_object_section("device_class"); | |
1359 | f->dump_int("id", i.first); | |
1360 | set<string> st; | |
1361 | calc_state_set(i.second, st); | |
1362 | for (auto& j : st) { | |
1363 | f->dump_string("flag", j); | |
1364 | } | |
1365 | f->close_section(); | |
1366 | } | |
1367 | f->close_section(); | |
f67539c2 TL |
1368 | f->open_object_section("stretch_mode"); |
1369 | { | |
1370 | f->dump_bool("change_stretch_mode", change_stretch_mode); | |
1371 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
1372 | f->dump_unsigned("new_stretch_bucket_count", new_stretch_bucket_count); | |
1373 | f->dump_unsigned("new_degraded_stretch_mode", new_degraded_stretch_mode); | |
1374 | f->dump_unsigned("new_recovering_stretch_mode", new_recovering_stretch_mode); | |
1375 | f->dump_int("new_stretch_mode_bucket", new_stretch_mode_bucket); | |
1376 | } | |
1377 | f->close_section(); | |
11fdf7f2 | 1378 | f->close_section(); |
7c673cae FG |
1379 | } |
1380 | ||
1381 | void OSDMap::Incremental::generate_test_instances(list<Incremental*>& o) | |
1382 | { | |
1383 | o.push_back(new Incremental); | |
1384 | } | |
1385 | ||
1386 | // ---------------------------------- | |
1387 | // OSDMap | |
1388 | ||
1389 | void OSDMap::set_epoch(epoch_t e) | |
1390 | { | |
1391 | epoch = e; | |
1392 | for (auto &pool : pools) | |
1393 | pool.second.last_change = e; | |
1394 | } | |
1395 | ||
33c7a0ef TL |
1396 | OSDMap::range_bits::range_bits() : ipv6(false) { |
1397 | memset(&bits, 0, sizeof(bits)); | |
1398 | } | |
1399 | ||
1400 | OSDMap::range_bits::range_bits(const entity_addr_t& addr) : ipv6(false) { | |
1401 | memset(&bits, 0, sizeof(bits)); | |
1402 | parse(addr); | |
1403 | } | |
1404 | ||
1405 | void OSDMap::range_bits::get_ipv6_bytes(unsigned const char *addr, | |
1406 | uint64_t *upper, uint64_t *lower) | |
1407 | { | |
1408 | *upper = ((uint64_t)(ntohl(*(uint32_t*)(addr)))) << 32 | | |
1409 | ((uint64_t)(ntohl(*(uint32_t*)(&addr[4])))); | |
1410 | *lower = ((uint64_t)(ntohl(*(uint32_t*)(&addr[8])))) << 32 | | |
1411 | ((uint64_t)(ntohl(*(uint32_t*)(&addr[12])))); | |
1412 | } | |
1413 | ||
1414 | void OSDMap::range_bits::parse(const entity_addr_t& addr) { | |
1415 | // parse it into meaningful data | |
1416 | if (addr.is_ipv6()) { | |
1417 | get_ipv6_bytes(addr.in6_addr().sin6_addr.s6_addr, | |
1418 | &bits.ipv6.upper_64_bits, &bits.ipv6.lower_64_bits); | |
1419 | int32_t lower_shift = std::min(128- | |
1420 | static_cast<int32_t>(addr.get_nonce()), 64); | |
1421 | int32_t upper_shift = std::max(64- //(128-b.first.get_nonce())-64 | |
1422 | static_cast<int32_t>(addr.get_nonce()), 0); | |
1423 | ||
1424 | auto get_mask = [](int32_t shift) -> uint64_t { | |
1425 | if (shift >= 0 && shift < 64) { | |
1426 | return UINT64_MAX << shift; | |
1427 | } | |
1428 | return 0; | |
1429 | }; | |
1430 | ||
1431 | bits.ipv6.lower_mask = get_mask(lower_shift); | |
1432 | bits.ipv6.upper_mask = get_mask(upper_shift); | |
1433 | ipv6 = true; | |
1434 | } else if (addr.is_ipv4()) { | |
1435 | bits.ipv4.ip_32_bits = ntohl(addr.in4_addr().sin_addr.s_addr); | |
1436 | if (addr.get_nonce() > 0) { | |
1437 | bits.ipv4.mask = UINT32_MAX << (32-addr.get_nonce()); | |
1438 | } else { | |
1439 | bits.ipv4.mask = 0; | |
1440 | } | |
1441 | } else { | |
1442 | // uh... | |
1443 | } | |
1444 | } | |
1445 | ||
1446 | bool OSDMap::range_bits::matches(const entity_addr_t& addr) const { | |
1447 | if (addr.is_ipv4() && !ipv6) { | |
1448 | return ((ntohl(addr.in4_addr().sin_addr.s_addr) & bits.ipv4.mask) == | |
1449 | (bits.ipv4.ip_32_bits & bits.ipv4.mask)); | |
1450 | } else if (addr.is_ipv6() && ipv6) { | |
1451 | uint64_t upper_64, lower_64; | |
1452 | get_ipv6_bytes(addr.in6_addr().sin6_addr.s6_addr, &upper_64, &lower_64); | |
1453 | return (((upper_64 & bits.ipv6.upper_mask) == | |
1454 | (bits.ipv6.upper_64_bits & bits.ipv6.upper_mask)) && | |
1455 | ((lower_64 & bits.ipv6.lower_mask) == | |
1456 | (bits.ipv6.lower_64_bits & bits.ipv6.lower_mask))); | |
1457 | } | |
1458 | return false; | |
1459 | } | |
1460 | ||
1461 | bool OSDMap::is_blocklisted(const entity_addr_t& orig, CephContext *cct) const | |
7c673cae | 1462 | { |
33c7a0ef TL |
1463 | if (cct) ldout(cct, 25) << "is_blocklisted: " << orig << dendl; |
1464 | if (blocklist.empty() && range_blocklist.empty()) { | |
1465 | if (cct) ldout(cct, 30) << "not blocklisted: " << orig << dendl; | |
7c673cae | 1466 | return false; |
11fdf7f2 TL |
1467 | } |
1468 | ||
f67539c2 | 1469 | // all blocklist entries are type ANY for nautilus+ |
11fdf7f2 TL |
1470 | // FIXME: avoid this copy! |
1471 | entity_addr_t a = orig; | |
9f95a23c | 1472 | if (require_osd_release < ceph_release_t::nautilus) { |
11fdf7f2 TL |
1473 | a.set_type(entity_addr_t::TYPE_LEGACY); |
1474 | } else { | |
1475 | a.set_type(entity_addr_t::TYPE_ANY); | |
1476 | } | |
7c673cae FG |
1477 | |
1478 | // this specific instance? | |
f67539c2 | 1479 | if (blocklist.count(a)) { |
33c7a0ef | 1480 | if (cct) ldout(cct, 20) << "blocklist contains " << a << dendl; |
7c673cae | 1481 | return true; |
11fdf7f2 | 1482 | } |
7c673cae | 1483 | |
f67539c2 | 1484 | // is entire ip blocklisted? |
7c673cae | 1485 | if (a.is_ip()) { |
11fdf7f2 TL |
1486 | a.set_port(0); |
1487 | a.set_nonce(0); | |
f67539c2 | 1488 | if (blocklist.count(a)) { |
33c7a0ef | 1489 | if (cct) ldout(cct, 20) << "blocklist contains " << a << dendl; |
11fdf7f2 TL |
1490 | return true; |
1491 | } | |
1492 | } | |
1493 | ||
33c7a0ef TL |
1494 | // is it in a blocklisted range? |
1495 | for (const auto& i : calculated_ranges) { | |
1496 | bool blocked = i.second.matches(a); | |
1497 | if (blocked) { | |
1498 | if (cct) ldout(cct, 20) << "range_blocklist contains " << a << dendl; | |
1499 | return true; | |
1500 | } | |
1501 | } | |
1502 | ||
1503 | if (cct) ldout(cct, 25) << "not blocklisted: " << orig << dendl; | |
11fdf7f2 TL |
1504 | return false; |
1505 | } | |
1506 | ||
33c7a0ef | 1507 | bool OSDMap::is_blocklisted(const entity_addrvec_t& av, CephContext *cct) const |
11fdf7f2 | 1508 | { |
33c7a0ef | 1509 | if (blocklist.empty() && range_blocklist.empty()) |
11fdf7f2 TL |
1510 | return false; |
1511 | ||
1512 | for (auto& a : av.v) { | |
33c7a0ef | 1513 | if (is_blocklisted(a, cct)) { |
7c673cae FG |
1514 | return true; |
1515 | } | |
1516 | } | |
1517 | ||
1518 | return false; | |
1519 | } | |
1520 | ||
33c7a0ef TL |
1521 | void OSDMap::get_blocklist(list<pair<entity_addr_t,utime_t> > *bl, |
1522 | std::list<std::pair<entity_addr_t,utime_t> > *rl) const | |
7c673cae | 1523 | { |
f67539c2 | 1524 | std::copy(blocklist.begin(), blocklist.end(), std::back_inserter(*bl)); |
33c7a0ef TL |
1525 | std::copy(range_blocklist.begin(), range_blocklist.end(), |
1526 | std::back_inserter(*rl)); | |
7c673cae FG |
1527 | } |
1528 | ||
33c7a0ef TL |
1529 | void OSDMap::get_blocklist(std::set<entity_addr_t> *bl, |
1530 | std::set<entity_addr_t> *rl) const | |
31f18b77 | 1531 | { |
f67539c2 | 1532 | for (const auto &i : blocklist) { |
31f18b77 FG |
1533 | bl->insert(i.first); |
1534 | } | |
33c7a0ef TL |
1535 | for (const auto &i : range_blocklist) { |
1536 | rl->insert(i.first); | |
1537 | } | |
31f18b77 FG |
1538 | } |
1539 | ||
7c673cae FG |
1540 | void OSDMap::set_max_osd(int m) |
1541 | { | |
7c673cae | 1542 | max_osd = m; |
f67539c2 TL |
1543 | osd_state.resize(max_osd, 0); |
1544 | osd_weight.resize(max_osd, CEPH_OSD_OUT); | |
1545 | osd_info.resize(max_osd); | |
1546 | osd_xinfo.resize(max_osd); | |
1547 | osd_addrs->client_addrs.resize(max_osd); | |
1548 | osd_addrs->cluster_addrs.resize(max_osd); | |
1549 | osd_addrs->hb_back_addrs.resize(max_osd); | |
1550 | osd_addrs->hb_front_addrs.resize(max_osd); | |
1551 | osd_uuid->resize(max_osd); | |
7c673cae | 1552 | if (osd_primary_affinity) |
f67539c2 | 1553 | osd_primary_affinity->resize(max_osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); |
7c673cae FG |
1554 | |
1555 | calc_num_osds(); | |
1556 | } | |
1557 | ||
1558 | int OSDMap::calc_num_osds() | |
1559 | { | |
1560 | num_osd = 0; | |
1561 | num_up_osd = 0; | |
1562 | num_in_osd = 0; | |
1563 | for (int i=0; i<max_osd; i++) { | |
1564 | if (osd_state[i] & CEPH_OSD_EXISTS) { | |
1565 | ++num_osd; | |
1566 | if (osd_state[i] & CEPH_OSD_UP) { | |
1567 | ++num_up_osd; | |
1568 | } | |
1569 | if (get_weight(i) != CEPH_OSD_OUT) { | |
1570 | ++num_in_osd; | |
1571 | } | |
1572 | } | |
1573 | } | |
1574 | return num_osd; | |
1575 | } | |
1576 | ||
3efd9988 FG |
1577 | void OSDMap::get_full_pools(CephContext *cct, |
1578 | set<int64_t> *full, | |
1579 | set<int64_t> *backfillfull, | |
1580 | set<int64_t> *nearfull) const | |
7c673cae | 1581 | { |
11fdf7f2 TL |
1582 | ceph_assert(full); |
1583 | ceph_assert(backfillfull); | |
1584 | ceph_assert(nearfull); | |
3efd9988 FG |
1585 | full->clear(); |
1586 | backfillfull->clear(); | |
1587 | nearfull->clear(); | |
1588 | ||
1589 | vector<int> full_osds; | |
1590 | vector<int> backfillfull_osds; | |
1591 | vector<int> nearfull_osds; | |
7c673cae FG |
1592 | for (int i = 0; i < max_osd; ++i) { |
1593 | if (exists(i) && is_up(i) && is_in(i)) { | |
1594 | if (osd_state[i] & CEPH_OSD_FULL) | |
3efd9988 | 1595 | full_osds.push_back(i); |
7c673cae | 1596 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) |
3efd9988 | 1597 | backfillfull_osds.push_back(i); |
7c673cae | 1598 | else if (osd_state[i] & CEPH_OSD_NEARFULL) |
3efd9988 | 1599 | nearfull_osds.push_back(i); |
7c673cae FG |
1600 | } |
1601 | } | |
3efd9988 FG |
1602 | |
1603 | for (auto i: full_osds) { | |
1604 | get_pool_ids_by_osd(cct, i, full); | |
1605 | } | |
1606 | for (auto i: backfillfull_osds) { | |
1607 | get_pool_ids_by_osd(cct, i, backfillfull); | |
1608 | } | |
1609 | for (auto i: nearfull_osds) { | |
1610 | get_pool_ids_by_osd(cct, i, nearfull); | |
1611 | } | |
7c673cae FG |
1612 | } |
1613 | ||
31f18b77 FG |
1614 | void OSDMap::get_full_osd_counts(set<int> *full, set<int> *backfill, |
1615 | set<int> *nearfull) const | |
1616 | { | |
1617 | full->clear(); | |
1618 | backfill->clear(); | |
1619 | nearfull->clear(); | |
1620 | for (int i = 0; i < max_osd; ++i) { | |
1621 | if (exists(i) && is_up(i) && is_in(i)) { | |
1622 | if (osd_state[i] & CEPH_OSD_FULL) | |
1623 | full->emplace(i); | |
1624 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1625 | backfill->emplace(i); | |
1626 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1627 | nearfull->emplace(i); | |
1628 | } | |
1629 | } | |
1630 | } | |
1631 | ||
7c673cae FG |
1632 | void OSDMap::get_all_osds(set<int32_t>& ls) const |
1633 | { | |
1634 | for (int i=0; i<max_osd; i++) | |
1635 | if (exists(i)) | |
1636 | ls.insert(i); | |
1637 | } | |
1638 | ||
1639 | void OSDMap::get_up_osds(set<int32_t>& ls) const | |
1640 | { | |
1641 | for (int i = 0; i < max_osd; i++) { | |
1642 | if (is_up(i)) | |
1643 | ls.insert(i); | |
1644 | } | |
1645 | } | |
1646 | ||
81eedcae | 1647 | void OSDMap::get_out_existing_osds(set<int32_t>& ls) const |
31f18b77 FG |
1648 | { |
1649 | for (int i = 0; i < max_osd; i++) { | |
81eedcae | 1650 | if (exists(i) && get_weight(i) == CEPH_OSD_OUT) |
31f18b77 FG |
1651 | ls.insert(i); |
1652 | } | |
1653 | } | |
1654 | ||
11fdf7f2 TL |
1655 | void OSDMap::get_flag_set(set<string> *flagset) const |
1656 | { | |
1657 | for (unsigned i = 0; i < sizeof(flags) * 8; ++i) { | |
1658 | if (flags & (1<<i)) { | |
1659 | flagset->insert(get_flag_string(flags & (1<<i))); | |
1660 | } | |
1661 | } | |
1662 | } | |
1663 | ||
7c673cae FG |
1664 | void OSDMap::calc_state_set(int state, set<string>& st) |
1665 | { | |
1666 | unsigned t = state; | |
1667 | for (unsigned s = 1; t; s <<= 1) { | |
1668 | if (t & s) { | |
1669 | t &= ~s; | |
1670 | st.insert(ceph_osd_state_name(s)); | |
1671 | } | |
1672 | } | |
1673 | } | |
1674 | ||
1675 | void OSDMap::adjust_osd_weights(const map<int,double>& weights, Incremental& inc) const | |
1676 | { | |
1677 | float max = 0; | |
1678 | for (const auto &weight : weights) { | |
1679 | if (weight.second > max) | |
1680 | max = weight.second; | |
1681 | } | |
1682 | ||
1683 | for (const auto &weight : weights) { | |
1684 | inc.new_weight[weight.first] = (unsigned)((weight.second / max) * CEPH_OSD_IN); | |
1685 | } | |
1686 | } | |
1687 | ||
1688 | int OSDMap::identify_osd(const entity_addr_t& addr) const | |
1689 | { | |
1690 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1691 | if (exists(i) && (get_addrs(i).contains(addr) || |
1692 | get_cluster_addrs(i).contains(addr))) | |
7c673cae FG |
1693 | return i; |
1694 | return -1; | |
1695 | } | |
1696 | ||
1697 | int OSDMap::identify_osd(const uuid_d& u) const | |
1698 | { | |
1699 | for (int i=0; i<max_osd; i++) | |
1700 | if (exists(i) && get_uuid(i) == u) | |
1701 | return i; | |
1702 | return -1; | |
1703 | } | |
1704 | ||
1705 | int OSDMap::identify_osd_on_all_channels(const entity_addr_t& addr) const | |
1706 | { | |
1707 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1708 | if (exists(i) && (get_addrs(i).contains(addr) || |
1709 | get_cluster_addrs(i).contains(addr) || | |
1710 | get_hb_back_addrs(i).contains(addr) || | |
1711 | get_hb_front_addrs(i).contains(addr))) | |
7c673cae FG |
1712 | return i; |
1713 | return -1; | |
1714 | } | |
1715 | ||
1716 | int OSDMap::find_osd_on_ip(const entity_addr_t& ip) const | |
1717 | { | |
1718 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1719 | if (exists(i) && (get_addrs(i).is_same_host(ip) || |
1720 | get_cluster_addrs(i).is_same_host(ip))) | |
7c673cae FG |
1721 | return i; |
1722 | return -1; | |
1723 | } | |
1724 | ||
1725 | ||
1726 | uint64_t OSDMap::get_features(int entity_type, uint64_t *pmask) const | |
1727 | { | |
1728 | uint64_t features = 0; // things we actually have | |
1729 | uint64_t mask = 0; // things we could have | |
1730 | ||
1731 | if (crush->has_nondefault_tunables()) | |
1732 | features |= CEPH_FEATURE_CRUSH_TUNABLES; | |
1733 | if (crush->has_nondefault_tunables2()) | |
1734 | features |= CEPH_FEATURE_CRUSH_TUNABLES2; | |
1735 | if (crush->has_nondefault_tunables3()) | |
1736 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1737 | if (crush->has_v4_buckets()) | |
1738 | features |= CEPH_FEATURE_CRUSH_V4; | |
1739 | if (crush->has_nondefault_tunables5()) | |
1740 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
c07f9fc5 FG |
1741 | if (crush->has_incompat_choose_args()) { |
1742 | features |= CEPH_FEATUREMASK_CRUSH_CHOOSE_ARGS; | |
1743 | } | |
7c673cae FG |
1744 | mask |= CEPH_FEATURES_CRUSH; |
1745 | ||
1e59de90 | 1746 | if (!pg_upmap.empty() || !pg_upmap_items.empty() || !pg_upmap_primaries.empty()) |
7c673cae FG |
1747 | features |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; |
1748 | mask |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1749 | ||
1750 | for (auto &pool: pools) { | |
1751 | if (pool.second.has_flag(pg_pool_t::FLAG_HASHPSPOOL)) { | |
1752 | features |= CEPH_FEATURE_OSDHASHPSPOOL; | |
1753 | } | |
7c673cae FG |
1754 | if (!pool.second.tiers.empty() || |
1755 | pool.second.is_tier()) { | |
1756 | features |= CEPH_FEATURE_OSD_CACHEPOOL; | |
1757 | } | |
20effc67 | 1758 | int ruleid = pool.second.get_crush_rule(); |
7c673cae FG |
1759 | if (ruleid >= 0) { |
1760 | if (crush->is_v2_rule(ruleid)) | |
1761 | features |= CEPH_FEATURE_CRUSH_V2; | |
1762 | if (crush->is_v3_rule(ruleid)) | |
1763 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1764 | if (crush->is_v5_rule(ruleid)) | |
1765 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
1766 | } | |
1767 | } | |
7c673cae | 1768 | mask |= CEPH_FEATURE_OSDHASHPSPOOL | CEPH_FEATURE_OSD_CACHEPOOL; |
7c673cae FG |
1769 | |
1770 | if (osd_primary_affinity) { | |
1771 | for (int i = 0; i < max_osd; ++i) { | |
1772 | if ((*osd_primary_affinity)[i] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
1773 | features |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1774 | break; | |
1775 | } | |
1776 | } | |
1777 | } | |
1778 | mask |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1779 | ||
1780 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1781 | const uint64_t jewel_features = CEPH_FEATURE_SERVER_JEWEL; | |
9f95a23c | 1782 | if (require_osd_release >= ceph_release_t::jewel) { |
7c673cae FG |
1783 | features |= jewel_features; |
1784 | } | |
1785 | mask |= jewel_features; | |
1786 | ||
1787 | const uint64_t kraken_features = CEPH_FEATUREMASK_SERVER_KRAKEN | |
1788 | | CEPH_FEATURE_MSG_ADDR2; | |
9f95a23c | 1789 | if (require_osd_release >= ceph_release_t::kraken) { |
7c673cae FG |
1790 | features |= kraken_features; |
1791 | } | |
1792 | mask |= kraken_features; | |
f67539c2 TL |
1793 | |
1794 | if (stretch_mode_enabled) { | |
1795 | features |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1796 | mask |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1797 | } | |
7c673cae FG |
1798 | } |
1799 | ||
9f95a23c | 1800 | if (require_min_compat_client >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
1801 | // if min_compat_client is >= nautilus, require v2 cephx signatures |
1802 | // from everyone | |
1803 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
9f95a23c | 1804 | } else if (require_osd_release >= ceph_release_t::nautilus && |
11fdf7f2 TL |
1805 | entity_type == CEPH_ENTITY_TYPE_OSD) { |
1806 | // if osds are >= nautilus, at least require the signatures from them | |
1807 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
1808 | } | |
1809 | mask |= CEPH_FEATUREMASK_CEPHX_V2; | |
1810 | ||
7c673cae FG |
1811 | if (pmask) |
1812 | *pmask = mask; | |
1813 | return features; | |
1814 | } | |
1815 | ||
9f95a23c | 1816 | ceph_release_t OSDMap::get_min_compat_client() const |
7c673cae FG |
1817 | { |
1818 | uint64_t f = get_features(CEPH_ENTITY_TYPE_CLIENT, nullptr); | |
1819 | ||
1820 | if (HAVE_FEATURE(f, OSDMAP_PG_UPMAP) || // v12.0.0-1733-g27d6f43 | |
31f18b77 | 1821 | HAVE_FEATURE(f, CRUSH_CHOOSE_ARGS)) { // v12.0.1-2172-gef1ef28 |
9f95a23c | 1822 | return ceph_release_t::luminous; // v12.2.0 |
7c673cae FG |
1823 | } |
1824 | if (HAVE_FEATURE(f, CRUSH_TUNABLES5)) { // v10.0.0-612-g043a737 | |
9f95a23c | 1825 | return ceph_release_t::jewel; // v10.2.0 |
7c673cae FG |
1826 | } |
1827 | if (HAVE_FEATURE(f, CRUSH_V4)) { // v0.91-678-g325fc56 | |
9f95a23c | 1828 | return ceph_release_t::hammer; // v0.94.0 |
7c673cae FG |
1829 | } |
1830 | if (HAVE_FEATURE(f, OSD_PRIMARY_AFFINITY) || // v0.76-553-gf825624 | |
1831 | HAVE_FEATURE(f, CRUSH_TUNABLES3) || // v0.76-395-ge20a55d | |
7c673cae | 1832 | HAVE_FEATURE(f, OSD_CACHEPOOL)) { // v0.67-401-gb91c1c5 |
9f95a23c | 1833 | return ceph_release_t::firefly; // v0.80.0 |
7c673cae FG |
1834 | } |
1835 | if (HAVE_FEATURE(f, CRUSH_TUNABLES2) || // v0.54-684-g0cc47ff | |
1836 | HAVE_FEATURE(f, OSDHASHPSPOOL)) { // v0.57-398-g8cc2b0f | |
9f95a23c | 1837 | return ceph_release_t::dumpling; // v0.67.0 |
7c673cae FG |
1838 | } |
1839 | if (HAVE_FEATURE(f, CRUSH_TUNABLES)) { // v0.48argonaut-206-g6f381af | |
9f95a23c | 1840 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae | 1841 | } |
9f95a23c | 1842 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae FG |
1843 | } |
1844 | ||
9f95a23c | 1845 | ceph_release_t OSDMap::get_require_min_compat_client() const |
11fdf7f2 TL |
1846 | { |
1847 | return require_min_compat_client; | |
1848 | } | |
1849 | ||
7c673cae FG |
1850 | void OSDMap::_calc_up_osd_features() |
1851 | { | |
1852 | bool first = true; | |
1853 | cached_up_osd_features = 0; | |
1854 | for (int osd = 0; osd < max_osd; ++osd) { | |
1855 | if (!is_up(osd)) | |
1856 | continue; | |
1857 | const osd_xinfo_t &xi = get_xinfo(osd); | |
3efd9988 FG |
1858 | if (xi.features == 0) |
1859 | continue; // bogus xinfo, maybe #20751 or similar, skipping | |
7c673cae FG |
1860 | if (first) { |
1861 | cached_up_osd_features = xi.features; | |
1862 | first = false; | |
1863 | } else { | |
1864 | cached_up_osd_features &= xi.features; | |
1865 | } | |
1866 | } | |
1867 | } | |
1868 | ||
1869 | uint64_t OSDMap::get_up_osd_features() const | |
1870 | { | |
1871 | return cached_up_osd_features; | |
1872 | } | |
1873 | ||
1874 | void OSDMap::dedup(const OSDMap *o, OSDMap *n) | |
1875 | { | |
11fdf7f2 | 1876 | using ceph::encode; |
7c673cae FG |
1877 | if (o->epoch == n->epoch) |
1878 | return; | |
1879 | ||
1880 | int diff = 0; | |
1881 | ||
1882 | // do addrs match? | |
1883 | if (o->max_osd != n->max_osd) | |
1884 | diff++; | |
1885 | for (int i = 0; i < o->max_osd && i < n->max_osd; i++) { | |
11fdf7f2 TL |
1886 | if ( n->osd_addrs->client_addrs[i] && o->osd_addrs->client_addrs[i] && |
1887 | *n->osd_addrs->client_addrs[i] == *o->osd_addrs->client_addrs[i]) | |
1888 | n->osd_addrs->client_addrs[i] = o->osd_addrs->client_addrs[i]; | |
7c673cae FG |
1889 | else |
1890 | diff++; | |
11fdf7f2 TL |
1891 | if ( n->osd_addrs->cluster_addrs[i] && o->osd_addrs->cluster_addrs[i] && |
1892 | *n->osd_addrs->cluster_addrs[i] == *o->osd_addrs->cluster_addrs[i]) | |
1893 | n->osd_addrs->cluster_addrs[i] = o->osd_addrs->cluster_addrs[i]; | |
7c673cae FG |
1894 | else |
1895 | diff++; | |
11fdf7f2 TL |
1896 | if ( n->osd_addrs->hb_back_addrs[i] && o->osd_addrs->hb_back_addrs[i] && |
1897 | *n->osd_addrs->hb_back_addrs[i] == *o->osd_addrs->hb_back_addrs[i]) | |
1898 | n->osd_addrs->hb_back_addrs[i] = o->osd_addrs->hb_back_addrs[i]; | |
7c673cae FG |
1899 | else |
1900 | diff++; | |
11fdf7f2 TL |
1901 | if ( n->osd_addrs->hb_front_addrs[i] && o->osd_addrs->hb_front_addrs[i] && |
1902 | *n->osd_addrs->hb_front_addrs[i] == *o->osd_addrs->hb_front_addrs[i]) | |
1903 | n->osd_addrs->hb_front_addrs[i] = o->osd_addrs->hb_front_addrs[i]; | |
7c673cae FG |
1904 | else |
1905 | diff++; | |
1906 | } | |
1907 | if (diff == 0) { | |
1908 | // zoinks, no differences at all! | |
1909 | n->osd_addrs = o->osd_addrs; | |
1910 | } | |
1911 | ||
1912 | // does crush match? | |
9f95a23c | 1913 | ceph::buffer::list oc, nc; |
11fdf7f2 TL |
1914 | encode(*o->crush, oc, CEPH_FEATURES_SUPPORTED_DEFAULT); |
1915 | encode(*n->crush, nc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
7c673cae FG |
1916 | if (oc.contents_equal(nc)) { |
1917 | n->crush = o->crush; | |
1918 | } | |
1919 | ||
1920 | // does pg_temp match? | |
31f18b77 FG |
1921 | if (*o->pg_temp == *n->pg_temp) |
1922 | n->pg_temp = o->pg_temp; | |
7c673cae FG |
1923 | |
1924 | // does primary_temp match? | |
1925 | if (o->primary_temp->size() == n->primary_temp->size()) { | |
1926 | if (*o->primary_temp == *n->primary_temp) | |
1927 | n->primary_temp = o->primary_temp; | |
1928 | } | |
1929 | ||
1930 | // do uuids match? | |
1931 | if (o->osd_uuid->size() == n->osd_uuid->size() && | |
1932 | *o->osd_uuid == *n->osd_uuid) | |
1933 | n->osd_uuid = o->osd_uuid; | |
1934 | } | |
1935 | ||
1936 | void OSDMap::clean_temps(CephContext *cct, | |
11fdf7f2 TL |
1937 | const OSDMap& oldmap, |
1938 | const OSDMap& nextmap, | |
1939 | Incremental *pending_inc) | |
7c673cae FG |
1940 | { |
1941 | ldout(cct, 10) << __func__ << dendl; | |
7c673cae | 1942 | |
11fdf7f2 | 1943 | for (auto pg : *nextmap.pg_temp) { |
7c673cae FG |
1944 | // if pool does not exist, remove any existing pg_temps associated with |
1945 | // it. we don't care about pg_temps on the pending_inc either; if there | |
1946 | // are new_pg_temp entries on the pending, clear them out just as well. | |
11fdf7f2 | 1947 | if (!nextmap.have_pg_pool(pg.first.pool())) { |
7c673cae FG |
1948 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first |
1949 | << " for nonexistent pool " << pg.first.pool() << dendl; | |
1950 | pending_inc->new_pg_temp[pg.first].clear(); | |
1951 | continue; | |
1952 | } | |
20effc67 TL |
1953 | if (!nextmap.pg_exists(pg.first)) { |
1954 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1955 | << " for nonexistent pg " << dendl; | |
1956 | pending_inc->new_pg_temp[pg.first].clear(); | |
1957 | continue; | |
1958 | } | |
7c673cae FG |
1959 | // all osds down? |
1960 | unsigned num_up = 0; | |
1961 | for (auto o : pg.second) { | |
11fdf7f2 | 1962 | if (!nextmap.is_down(o)) { |
7c673cae FG |
1963 | ++num_up; |
1964 | break; | |
1965 | } | |
1966 | } | |
1967 | if (num_up == 0) { | |
1968 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1969 | << " with all down osds" << pg.second << dendl; | |
1970 | pending_inc->new_pg_temp[pg.first].clear(); | |
1971 | continue; | |
1972 | } | |
1973 | // redundant pg_temp? | |
1974 | vector<int> raw_up; | |
1975 | int primary; | |
11fdf7f2 | 1976 | nextmap.pg_to_raw_up(pg.first, &raw_up, &primary); |
91327a77 | 1977 | bool remove = false; |
11fdf7f2 | 1978 | if (raw_up == pg.second) { |
7c673cae FG |
1979 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1980 | << pg.second << " that matches raw_up mapping" << dendl; | |
91327a77 AA |
1981 | remove = true; |
1982 | } | |
1983 | // oversized pg_temp? | |
11fdf7f2 | 1984 | if (pg.second.size() > nextmap.get_pg_pool(pg.first.pool())->get_size()) { |
91327a77 AA |
1985 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1986 | << pg.second << " exceeds pool size" << dendl; | |
1987 | remove = true; | |
1988 | } | |
1989 | if (remove) { | |
11fdf7f2 | 1990 | if (oldmap.pg_temp->count(pg.first)) |
7c673cae FG |
1991 | pending_inc->new_pg_temp[pg.first].clear(); |
1992 | else | |
1993 | pending_inc->new_pg_temp.erase(pg.first); | |
1994 | } | |
1995 | } | |
1996 | ||
11fdf7f2 | 1997 | for (auto &pg : *nextmap.primary_temp) { |
7c673cae | 1998 | // primary down? |
11fdf7f2 | 1999 | if (nextmap.is_down(pg.second)) { |
7c673cae FG |
2000 | ldout(cct, 10) << __func__ << " removing primary_temp " << pg.first |
2001 | << " to down " << pg.second << dendl; | |
2002 | pending_inc->new_primary_temp[pg.first] = -1; | |
2003 | continue; | |
2004 | } | |
2005 | // redundant primary_temp? | |
2006 | vector<int> real_up, templess_up; | |
2007 | int real_primary, templess_primary; | |
2008 | pg_t pgid = pg.first; | |
11fdf7f2 TL |
2009 | nextmap.pg_to_acting_osds(pgid, &real_up, &real_primary); |
2010 | nextmap.pg_to_raw_up(pgid, &templess_up, &templess_primary); | |
7c673cae FG |
2011 | if (real_primary == templess_primary){ |
2012 | ldout(cct, 10) << __func__ << " removing primary_temp " | |
2013 | << pgid << " -> " << real_primary | |
2014 | << " (unnecessary/redundant)" << dendl; | |
11fdf7f2 | 2015 | if (oldmap.primary_temp->count(pgid)) |
7c673cae FG |
2016 | pending_inc->new_primary_temp[pgid] = -1; |
2017 | else | |
2018 | pending_inc->new_primary_temp.erase(pgid); | |
2019 | } | |
2020 | } | |
2021 | } | |
2022 | ||
494da23a | 2023 | void OSDMap::get_upmap_pgs(vector<pg_t> *upmap_pgs) const |
94b18763 | 2024 | { |
494da23a TL |
2025 | upmap_pgs->reserve(pg_upmap.size() + pg_upmap_items.size()); |
2026 | for (auto& p : pg_upmap) | |
2027 | upmap_pgs->push_back(p.first); | |
2028 | for (auto& p : pg_upmap_items) | |
2029 | upmap_pgs->push_back(p.first); | |
2030 | } | |
94b18763 | 2031 | |
494da23a TL |
2032 | bool OSDMap::check_pg_upmaps( |
2033 | CephContext *cct, | |
2034 | const vector<pg_t>& to_check, | |
2035 | vector<pg_t> *to_cancel, | |
2036 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> *to_remap) const | |
2037 | { | |
2038 | bool any_change = false; | |
2039 | map<int, map<int, float>> rule_weight_map; | |
28e407b8 | 2040 | for (auto& pg : to_check) { |
494da23a | 2041 | const pg_pool_t *pi = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2042 | if (!pi || pg.ps() >= pi->get_pg_num_pending()) { |
2043 | ldout(cct, 0) << __func__ << " pg " << pg << " is gone or merge source" | |
2044 | << dendl; | |
494da23a | 2045 | to_cancel->push_back(pg); |
11fdf7f2 TL |
2046 | continue; |
2047 | } | |
2048 | if (pi->is_pending_merge(pg, nullptr)) { | |
2049 | ldout(cct, 0) << __func__ << " pg " << pg << " is pending merge" | |
2050 | << dendl; | |
494da23a | 2051 | to_cancel->push_back(pg); |
94b18763 FG |
2052 | continue; |
2053 | } | |
494da23a TL |
2054 | vector<int> raw, up; |
2055 | pg_to_raw_upmap(pg, &raw, &up); | |
494da23a TL |
2056 | auto crush_rule = get_pg_pool_crush_rule(pg); |
2057 | auto r = crush->verify_upmap(cct, | |
2058 | crush_rule, | |
2059 | get_pg_pool_size(pg), | |
2060 | up); | |
a8e16298 TL |
2061 | if (r < 0) { |
2062 | ldout(cct, 0) << __func__ << " verify_upmap of pg " << pg | |
2063 | << " returning " << r | |
2064 | << dendl; | |
494da23a | 2065 | to_cancel->push_back(pg); |
a8e16298 TL |
2066 | continue; |
2067 | } | |
2068 | // below we check against crush-topology changing.. | |
28e407b8 AA |
2069 | map<int, float> weight_map; |
2070 | auto it = rule_weight_map.find(crush_rule); | |
2071 | if (it == rule_weight_map.end()) { | |
494da23a | 2072 | auto r = crush->get_rule_weight_osd_map(crush_rule, &weight_map); |
28e407b8 AA |
2073 | if (r < 0) { |
2074 | lderr(cct) << __func__ << " unable to get crush weight_map for " | |
494da23a TL |
2075 | << "crush_rule " << crush_rule |
2076 | << dendl; | |
28e407b8 AA |
2077 | continue; |
2078 | } | |
2079 | rule_weight_map[crush_rule] = weight_map; | |
2080 | } else { | |
2081 | weight_map = it->second; | |
2082 | } | |
28e407b8 | 2083 | ldout(cct, 10) << __func__ << " pg " << pg |
28e407b8 | 2084 | << " weight_map " << weight_map |
94b18763 | 2085 | << dendl; |
a8e16298 | 2086 | for (auto osd : up) { |
28e407b8 AA |
2087 | auto it = weight_map.find(osd); |
2088 | if (it == weight_map.end()) { | |
92f5a8d4 TL |
2089 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd << " is gone or has " |
2090 | << "been moved out of the specific crush-tree" | |
2091 | << dendl; | |
494da23a | 2092 | to_cancel->push_back(pg); |
94b18763 FG |
2093 | break; |
2094 | } | |
494da23a | 2095 | auto adjusted_weight = get_weightf(it->first) * it->second; |
28e407b8 | 2096 | if (adjusted_weight == 0) { |
92f5a8d4 TL |
2097 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd |
2098 | << " is out/crush-out" | |
2099 | << dendl; | |
494da23a | 2100 | to_cancel->push_back(pg); |
94b18763 FG |
2101 | break; |
2102 | } | |
2103 | } | |
eafe8130 TL |
2104 | if (!to_cancel->empty() && to_cancel->back() == pg) |
2105 | continue; | |
2106 | // okay, upmap is valid | |
2107 | // continue to check if it is still necessary | |
2108 | auto i = pg_upmap.find(pg); | |
a4b75251 TL |
2109 | if (i != pg_upmap.end()) { |
2110 | if (i->second == raw) { | |
1e59de90 | 2111 | ldout(cct, 10) << __func__ << "removing redundant pg_upmap " << i->first << " " |
a4b75251 TL |
2112 | << i->second << dendl; |
2113 | to_cancel->push_back(pg); | |
2114 | continue; | |
2115 | } | |
2116 | if ((int)i->second.size() != get_pg_pool_size(pg)) { | |
1e59de90 | 2117 | ldout(cct, 10) << __func__ << "removing pg_upmap " << i->first << " " |
a4b75251 TL |
2118 | << i->second << " != pool size " << get_pg_pool_size(pg) |
2119 | << dendl; | |
2120 | to_cancel->push_back(pg); | |
2121 | continue; | |
2122 | } | |
eafe8130 TL |
2123 | } |
2124 | auto j = pg_upmap_items.find(pg); | |
2125 | if (j != pg_upmap_items.end()) { | |
2126 | mempool::osdmap::vector<pair<int,int>> newmap; | |
2127 | for (auto& p : j->second) { | |
1e59de90 TL |
2128 | auto osd_from = p.first; |
2129 | auto osd_to = p.second; | |
2130 | if (std::find(raw.begin(), raw.end(), osd_from) == raw.end()) { | |
eafe8130 | 2131 | // cancel mapping if source osd does not exist anymore |
1e59de90 | 2132 | ldout(cct, 20) << __func__ << " pg_upmap_items (source osd does not exist) " << pg_upmap_items << dendl; |
eafe8130 TL |
2133 | continue; |
2134 | } | |
1e59de90 TL |
2135 | if (osd_to != CRUSH_ITEM_NONE && osd_to < max_osd && |
2136 | osd_to >= 0 && osd_weight[osd_to] == 0) { | |
eafe8130 | 2137 | // cancel mapping if target osd is out |
1e59de90 | 2138 | ldout(cct, 20) << __func__ << " pg_upmap_items (target osd is out) " << pg_upmap_items << dendl; |
eafe8130 TL |
2139 | continue; |
2140 | } | |
2141 | newmap.push_back(p); | |
2142 | } | |
2143 | if (newmap.empty()) { | |
1e59de90 | 2144 | ldout(cct, 10) << __func__ << " removing no-op pg_upmap_items " |
eafe8130 TL |
2145 | << j->first << " " << j->second |
2146 | << dendl; | |
2147 | to_cancel->push_back(pg); | |
aee94f69 TL |
2148 | } else if (newmap != j->second) { |
2149 | // check partial no-op here. | |
1e59de90 | 2150 | ldout(cct, 10) << __func__ << " simplifying partially no-op pg_upmap_items " |
eafe8130 TL |
2151 | << j->first << " " << j->second |
2152 | << " -> " << newmap | |
2153 | << dendl; | |
2154 | to_remap->insert({pg, newmap}); | |
2155 | any_change = true; | |
2156 | } | |
2157 | } | |
28e407b8 | 2158 | } |
494da23a TL |
2159 | any_change = any_change || !to_cancel->empty(); |
2160 | return any_change; | |
2161 | } | |
2162 | ||
2163 | void OSDMap::clean_pg_upmaps( | |
2164 | CephContext *cct, | |
2165 | Incremental *pending_inc, | |
2166 | const vector<pg_t>& to_cancel, | |
2167 | const map<pg_t, mempool::osdmap::vector<pair<int,int>>>& to_remap) const | |
2168 | { | |
28e407b8 | 2169 | for (auto &pg: to_cancel) { |
494da23a TL |
2170 | auto i = pending_inc->new_pg_upmap.find(pg); |
2171 | if (i != pending_inc->new_pg_upmap.end()) { | |
2172 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2173 | << "pg_upmap entry " | |
2174 | << i->first << "->" << i->second | |
2175 | << dendl; | |
2176 | pending_inc->new_pg_upmap.erase(i); | |
94b18763 | 2177 | } |
494da23a TL |
2178 | auto j = pg_upmap.find(pg); |
2179 | if (j != pg_upmap.end()) { | |
2180 | ldout(cct, 10) << __func__ << " cancel invalid pg_upmap entry " | |
2181 | << j->first << "->" << j->second | |
2182 | << dendl; | |
2183 | pending_inc->old_pg_upmap.insert(pg); | |
2184 | } | |
2185 | auto p = pending_inc->new_pg_upmap_items.find(pg); | |
2186 | if (p != pending_inc->new_pg_upmap_items.end()) { | |
2187 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2188 | << "pg_upmap_items entry " | |
2189 | << p->first << "->" << p->second | |
2190 | << dendl; | |
2191 | pending_inc->new_pg_upmap_items.erase(p); | |
2192 | } | |
2193 | auto q = pg_upmap_items.find(pg); | |
2194 | if (q != pg_upmap_items.end()) { | |
2195 | ldout(cct, 10) << __func__ << " cancel invalid " | |
2196 | << "pg_upmap_items entry " | |
2197 | << q->first << "->" << q->second | |
2198 | << dendl; | |
2199 | pending_inc->old_pg_upmap_items.insert(pg); | |
94b18763 FG |
2200 | } |
2201 | } | |
494da23a TL |
2202 | for (auto& i : to_remap) |
2203 | pending_inc->new_pg_upmap_items[i.first] = i.second; | |
2204 | } | |
2205 | ||
2206 | bool OSDMap::clean_pg_upmaps( | |
2207 | CephContext *cct, | |
2208 | Incremental *pending_inc) const | |
2209 | { | |
2210 | ldout(cct, 10) << __func__ << dendl; | |
2211 | vector<pg_t> to_check; | |
2212 | vector<pg_t> to_cancel; | |
2213 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> to_remap; | |
2214 | ||
2215 | get_upmap_pgs(&to_check); | |
2216 | auto any_change = check_pg_upmaps(cct, to_check, &to_cancel, &to_remap); | |
2217 | clean_pg_upmaps(cct, pending_inc, to_cancel, to_remap); | |
1e59de90 TL |
2218 | //TODO: Create these 3 functions for pg_upmap_primaries and so they can be checked |
2219 | // and cleaned in the same way as pg_upmap. This is not critical since invalid | |
2220 | // pg_upmap_primaries are never applied, (the final check is in _apply_upmap). | |
494da23a | 2221 | return any_change; |
94b18763 FG |
2222 | } |
2223 | ||
7c673cae FG |
2224 | int OSDMap::apply_incremental(const Incremental &inc) |
2225 | { | |
f67539c2 | 2226 | new_blocklist_entries = false; |
7c673cae FG |
2227 | if (inc.epoch == 1) |
2228 | fsid = inc.fsid; | |
2229 | else if (inc.fsid != fsid) | |
2230 | return -EINVAL; | |
2231 | ||
11fdf7f2 | 2232 | ceph_assert(inc.epoch == epoch+1); |
7c673cae FG |
2233 | |
2234 | epoch++; | |
2235 | modified = inc.modified; | |
2236 | ||
2237 | // full map? | |
2238 | if (inc.fullmap.length()) { | |
9f95a23c | 2239 | ceph::buffer::list bl(inc.fullmap); |
7c673cae FG |
2240 | decode(bl); |
2241 | return 0; | |
2242 | } | |
2243 | ||
2244 | // nope, incremental. | |
31f18b77 | 2245 | if (inc.new_flags >= 0) { |
7c673cae | 2246 | flags = inc.new_flags; |
31f18b77 FG |
2247 | // the below is just to cover a newly-upgraded luminous mon |
2248 | // cluster that has to set require_jewel_osds or | |
2249 | // require_kraken_osds before the osds can be upgraded to | |
2250 | // luminous. | |
2251 | if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { | |
9f95a23c TL |
2252 | if (require_osd_release < ceph_release_t::kraken) { |
2253 | require_osd_release = ceph_release_t::kraken; | |
31f18b77 FG |
2254 | } |
2255 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
9f95a23c TL |
2256 | if (require_osd_release < ceph_release_t::jewel) { |
2257 | require_osd_release = ceph_release_t::jewel; | |
31f18b77 FG |
2258 | } |
2259 | } | |
2260 | } | |
7c673cae FG |
2261 | |
2262 | if (inc.new_max_osd >= 0) | |
2263 | set_max_osd(inc.new_max_osd); | |
2264 | ||
2265 | if (inc.new_pool_max != -1) | |
2266 | pool_max = inc.new_pool_max; | |
2267 | ||
2268 | for (const auto &pool : inc.new_pools) { | |
2269 | pools[pool.first] = pool.second; | |
2270 | pools[pool.first].last_change = epoch; | |
2271 | } | |
2272 | ||
11fdf7f2 TL |
2273 | new_removed_snaps = inc.new_removed_snaps; |
2274 | new_purged_snaps = inc.new_purged_snaps; | |
2275 | for (auto p = new_removed_snaps.begin(); | |
2276 | p != new_removed_snaps.end(); | |
2277 | ++p) { | |
2278 | removed_snaps_queue[p->first].union_of(p->second); | |
2279 | } | |
2280 | for (auto p = new_purged_snaps.begin(); | |
2281 | p != new_purged_snaps.end(); | |
2282 | ++p) { | |
2283 | auto q = removed_snaps_queue.find(p->first); | |
2284 | ceph_assert(q != removed_snaps_queue.end()); | |
2285 | q->second.subtract(p->second); | |
2286 | if (q->second.empty()) { | |
2287 | removed_snaps_queue.erase(q); | |
2288 | } | |
2289 | } | |
2290 | ||
2291 | if (inc.new_last_up_change != utime_t()) { | |
2292 | last_up_change = inc.new_last_up_change; | |
2293 | } | |
2294 | if (inc.new_last_in_change != utime_t()) { | |
2295 | last_in_change = inc.new_last_in_change; | |
2296 | } | |
2297 | ||
7c673cae FG |
2298 | for (const auto &pname : inc.new_pool_names) { |
2299 | auto pool_name_entry = pool_name.find(pname.first); | |
2300 | if (pool_name_entry != pool_name.end()) { | |
2301 | name_pool.erase(pool_name_entry->second); | |
2302 | pool_name_entry->second = pname.second; | |
2303 | } else { | |
2304 | pool_name[pname.first] = pname.second; | |
2305 | } | |
2306 | name_pool[pname.second] = pname.first; | |
2307 | } | |
2308 | ||
2309 | for (const auto &pool : inc.old_pools) { | |
2310 | pools.erase(pool); | |
2311 | name_pool.erase(pool_name[pool]); | |
2312 | pool_name.erase(pool); | |
2313 | } | |
2314 | ||
2315 | for (const auto &weight : inc.new_weight) { | |
2316 | set_weight(weight.first, weight.second); | |
2317 | ||
2318 | // if we are marking in, clear the AUTOOUT and NEW bits, and clear | |
2319 | // xinfo old_weight. | |
2320 | if (weight.second) { | |
2321 | osd_state[weight.first] &= ~(CEPH_OSD_AUTOOUT | CEPH_OSD_NEW); | |
2322 | osd_xinfo[weight.first].old_weight = 0; | |
2323 | } | |
2324 | } | |
2325 | ||
2326 | for (const auto &primary_affinity : inc.new_primary_affinity) { | |
2327 | set_primary_affinity(primary_affinity.first, primary_affinity.second); | |
2328 | } | |
2329 | ||
2330 | // erasure_code_profiles | |
2331 | for (const auto &profile : inc.old_erasure_code_profiles) | |
2332 | erasure_code_profiles.erase(profile); | |
2333 | ||
2334 | for (const auto &profile : inc.new_erasure_code_profiles) { | |
2335 | set_erasure_code_profile(profile.first, profile.second); | |
2336 | } | |
2337 | ||
2338 | // up/down | |
2339 | for (const auto &state : inc.new_state) { | |
2340 | const auto osd = state.first; | |
2341 | int s = state.second ? state.second : CEPH_OSD_UP; | |
2342 | if ((osd_state[osd] & CEPH_OSD_UP) && | |
2343 | (s & CEPH_OSD_UP)) { | |
2344 | osd_info[osd].down_at = epoch; | |
2345 | osd_xinfo[osd].down_stamp = modified; | |
2346 | } | |
2347 | if ((osd_state[osd] & CEPH_OSD_EXISTS) && | |
2348 | (s & CEPH_OSD_EXISTS)) { | |
2349 | // osd is destroyed; clear out anything interesting. | |
2350 | (*osd_uuid)[osd] = uuid_d(); | |
2351 | osd_info[osd] = osd_info_t(); | |
2352 | osd_xinfo[osd] = osd_xinfo_t(); | |
2353 | set_primary_affinity(osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
11fdf7f2 TL |
2354 | osd_addrs->client_addrs[osd].reset(new entity_addrvec_t()); |
2355 | osd_addrs->cluster_addrs[osd].reset(new entity_addrvec_t()); | |
2356 | osd_addrs->hb_front_addrs[osd].reset(new entity_addrvec_t()); | |
2357 | osd_addrs->hb_back_addrs[osd].reset(new entity_addrvec_t()); | |
7c673cae FG |
2358 | osd_state[osd] = 0; |
2359 | } else { | |
2360 | osd_state[osd] ^= s; | |
2361 | } | |
2362 | } | |
2363 | ||
2364 | for (const auto &client : inc.new_up_client) { | |
2365 | osd_state[client.first] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; | |
9f95a23c | 2366 | osd_state[client.first] &= ~CEPH_OSD_STOP; // if any |
11fdf7f2 TL |
2367 | osd_addrs->client_addrs[client.first].reset( |
2368 | new entity_addrvec_t(client.second)); | |
2369 | osd_addrs->hb_back_addrs[client.first].reset( | |
2370 | new entity_addrvec_t(inc.new_hb_back_up.find(client.first)->second)); | |
2371 | osd_addrs->hb_front_addrs[client.first].reset( | |
2372 | new entity_addrvec_t(inc.new_hb_front_up.find(client.first)->second)); | |
7c673cae FG |
2373 | |
2374 | osd_info[client.first].up_from = epoch; | |
2375 | } | |
2376 | ||
2377 | for (const auto &cluster : inc.new_up_cluster) | |
11fdf7f2 TL |
2378 | osd_addrs->cluster_addrs[cluster.first].reset( |
2379 | new entity_addrvec_t(cluster.second)); | |
7c673cae FG |
2380 | |
2381 | // info | |
2382 | for (const auto &thru : inc.new_up_thru) | |
2383 | osd_info[thru.first].up_thru = thru.second; | |
2384 | ||
2385 | for (const auto &interval : inc.new_last_clean_interval) { | |
2386 | osd_info[interval.first].last_clean_begin = interval.second.first; | |
2387 | osd_info[interval.first].last_clean_end = interval.second.second; | |
2388 | } | |
2389 | ||
2390 | for (const auto &lost : inc.new_lost) | |
2391 | osd_info[lost.first].lost_at = lost.second; | |
2392 | ||
2393 | // xinfo | |
2394 | for (const auto &xinfo : inc.new_xinfo) | |
2395 | osd_xinfo[xinfo.first] = xinfo.second; | |
2396 | ||
2397 | // uuid | |
2398 | for (const auto &uuid : inc.new_uuid) | |
2399 | (*osd_uuid)[uuid.first] = uuid.second; | |
2400 | ||
2401 | // pg rebuild | |
2402 | for (const auto &pg : inc.new_pg_temp) { | |
2403 | if (pg.second.empty()) | |
2404 | pg_temp->erase(pg.first); | |
2405 | else | |
31f18b77 FG |
2406 | pg_temp->set(pg.first, pg.second); |
2407 | } | |
2408 | if (!inc.new_pg_temp.empty()) { | |
2409 | // make sure pg_temp is efficiently stored | |
2410 | pg_temp->rebuild(); | |
7c673cae FG |
2411 | } |
2412 | ||
2413 | for (const auto &pg : inc.new_primary_temp) { | |
2414 | if (pg.second == -1) | |
2415 | primary_temp->erase(pg.first); | |
2416 | else | |
2417 | (*primary_temp)[pg.first] = pg.second; | |
2418 | } | |
2419 | ||
2420 | for (auto& p : inc.new_pg_upmap) { | |
2421 | pg_upmap[p.first] = p.second; | |
2422 | } | |
2423 | for (auto& pg : inc.old_pg_upmap) { | |
2424 | pg_upmap.erase(pg); | |
2425 | } | |
2426 | for (auto& p : inc.new_pg_upmap_items) { | |
2427 | pg_upmap_items[p.first] = p.second; | |
2428 | } | |
2429 | for (auto& pg : inc.old_pg_upmap_items) { | |
2430 | pg_upmap_items.erase(pg); | |
2431 | } | |
2432 | ||
1e59de90 TL |
2433 | for (auto& [pg, prim] : inc.new_pg_upmap_primary) { |
2434 | pg_upmap_primaries[pg] = prim; | |
2435 | } | |
2436 | for (auto& pg : inc.old_pg_upmap_primary) { | |
2437 | pg_upmap_primaries.erase(pg); | |
2438 | } | |
2439 | ||
f67539c2 TL |
2440 | // blocklist |
2441 | if (!inc.new_blocklist.empty()) { | |
2442 | blocklist.insert(inc.new_blocklist.begin(),inc.new_blocklist.end()); | |
2443 | new_blocklist_entries = true; | |
7c673cae | 2444 | } |
f67539c2 TL |
2445 | for (const auto &addr : inc.old_blocklist) |
2446 | blocklist.erase(addr); | |
7c673cae | 2447 | |
33c7a0ef TL |
2448 | for (const auto& addr_p : inc.new_range_blocklist) { |
2449 | range_blocklist.insert(addr_p); | |
2450 | calculated_ranges.emplace(addr_p.first, addr_p.first); | |
2451 | new_blocklist_entries = true; | |
2452 | } | |
2453 | for (const auto &addr : inc.old_range_blocklist) { | |
2454 | calculated_ranges.erase(addr); | |
2455 | range_blocklist.erase(addr); | |
2456 | } | |
2457 | ||
81eedcae TL |
2458 | for (auto& i : inc.new_crush_node_flags) { |
2459 | if (i.second) { | |
2460 | crush_node_flags[i.first] = i.second; | |
2461 | } else { | |
2462 | crush_node_flags.erase(i.first); | |
2463 | } | |
2464 | } | |
2465 | ||
2466 | for (auto& i : inc.new_device_class_flags) { | |
2467 | if (i.second) { | |
2468 | device_class_flags[i.first] = i.second; | |
2469 | } else { | |
2470 | device_class_flags.erase(i.first); | |
2471 | } | |
2472 | } | |
2473 | ||
7c673cae FG |
2474 | // cluster snapshot? |
2475 | if (inc.cluster_snapshot.length()) { | |
2476 | cluster_snapshot = inc.cluster_snapshot; | |
2477 | cluster_snapshot_epoch = inc.epoch; | |
2478 | } else { | |
2479 | cluster_snapshot.clear(); | |
2480 | cluster_snapshot_epoch = 0; | |
2481 | } | |
2482 | ||
2483 | if (inc.new_nearfull_ratio >= 0) { | |
2484 | nearfull_ratio = inc.new_nearfull_ratio; | |
2485 | } | |
2486 | if (inc.new_backfillfull_ratio >= 0) { | |
2487 | backfillfull_ratio = inc.new_backfillfull_ratio; | |
2488 | } | |
2489 | if (inc.new_full_ratio >= 0) { | |
2490 | full_ratio = inc.new_full_ratio; | |
2491 | } | |
9f95a23c | 2492 | if (inc.new_require_min_compat_client > ceph_release_t::unknown) { |
7c673cae FG |
2493 | require_min_compat_client = inc.new_require_min_compat_client; |
2494 | } | |
9f95a23c | 2495 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
31f18b77 | 2496 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2497 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 2498 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 2499 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2500 | } |
2501 | } | |
7c673cae | 2502 | |
9f95a23c | 2503 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
11fdf7f2 | 2504 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2505 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
2506 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
2507 | } | |
2508 | } | |
7c673cae FG |
2509 | // do new crush map last (after up/down stuff) |
2510 | if (inc.crush.length()) { | |
9f95a23c | 2511 | ceph::buffer::list bl(inc.crush); |
11fdf7f2 | 2512 | auto blp = bl.cbegin(); |
7c673cae FG |
2513 | crush.reset(new CrushWrapper); |
2514 | crush->decode(blp); | |
9f95a23c | 2515 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 FG |
2516 | // only increment if this is a luminous-encoded osdmap, lest |
2517 | // the mon's crush_version diverge from what the osds or others | |
2518 | // are decoding and applying on their end. if we won't encode | |
2519 | // it in the canonical version, don't change it. | |
2520 | ++crush_version; | |
2521 | } | |
81eedcae TL |
2522 | for (auto it = device_class_flags.begin(); |
2523 | it != device_class_flags.end();) { | |
2524 | const char* class_name = crush->get_class_name(it->first); | |
2525 | if (!class_name) // device class is gone | |
2526 | it = device_class_flags.erase(it); | |
2527 | else | |
2528 | it++; | |
2529 | } | |
7c673cae FG |
2530 | } |
2531 | ||
f67539c2 TL |
2532 | if (inc.change_stretch_mode) { |
2533 | stretch_mode_enabled = inc.stretch_mode_enabled; | |
2534 | stretch_bucket_count = inc.new_stretch_bucket_count; | |
2535 | degraded_stretch_mode = inc.new_degraded_stretch_mode; | |
2536 | recovering_stretch_mode = inc.new_recovering_stretch_mode; | |
2537 | stretch_mode_bucket = inc.new_stretch_mode_bucket; | |
2538 | } | |
2539 | ||
1e59de90 TL |
2540 | switch (inc.mutate_allow_crimson) { |
2541 | case Incremental::mutate_allow_crimson_t::NONE: | |
2542 | break; | |
2543 | case Incremental::mutate_allow_crimson_t::SET: | |
2544 | allow_crimson = true; | |
2545 | break; | |
2546 | case Incremental::mutate_allow_crimson_t::CLEAR: | |
2547 | allow_crimson = false; | |
2548 | break; | |
2549 | } | |
2550 | ||
7c673cae FG |
2551 | calc_num_osds(); |
2552 | _calc_up_osd_features(); | |
2553 | return 0; | |
2554 | } | |
2555 | ||
2556 | // mapping | |
2557 | int OSDMap::map_to_pg( | |
2558 | int64_t poolid, | |
2559 | const string& name, | |
2560 | const string& key, | |
2561 | const string& nspace, | |
2562 | pg_t *pg) const | |
2563 | { | |
2564 | // calculate ps (placement seed) | |
2565 | const pg_pool_t *pool = get_pg_pool(poolid); | |
2566 | if (!pool) | |
2567 | return -ENOENT; | |
2568 | ps_t ps; | |
2569 | if (!key.empty()) | |
2570 | ps = pool->hash_key(key, nspace); | |
2571 | else | |
2572 | ps = pool->hash_key(name, nspace); | |
2573 | *pg = pg_t(ps, poolid); | |
2574 | return 0; | |
2575 | } | |
2576 | ||
2577 | int OSDMap::object_locator_to_pg( | |
2578 | const object_t& oid, const object_locator_t& loc, pg_t &pg) const | |
2579 | { | |
2580 | if (loc.hash >= 0) { | |
2581 | if (!get_pg_pool(loc.get_pool())) { | |
2582 | return -ENOENT; | |
2583 | } | |
2584 | pg = pg_t(loc.hash, loc.get_pool()); | |
2585 | return 0; | |
2586 | } | |
2587 | return map_to_pg(loc.get_pool(), oid.name, loc.key, loc.nspace, &pg); | |
2588 | } | |
2589 | ||
2590 | ceph_object_layout OSDMap::make_object_layout( | |
2591 | object_t oid, int pg_pool, string nspace) const | |
2592 | { | |
2593 | object_locator_t loc(pg_pool, nspace); | |
2594 | ||
2595 | ceph_object_layout ol; | |
2596 | pg_t pgid = object_locator_to_pg(oid, loc); | |
2597 | ol.ol_pgid = pgid.get_old_pg().v; | |
2598 | ol.ol_stripe_unit = 0; | |
2599 | return ol; | |
2600 | } | |
2601 | ||
2602 | void OSDMap::_remove_nonexistent_osds(const pg_pool_t& pool, | |
2603 | vector<int>& osds) const | |
2604 | { | |
2605 | if (pool.can_shift_osds()) { | |
2606 | unsigned removed = 0; | |
2607 | for (unsigned i = 0; i < osds.size(); i++) { | |
2608 | if (!exists(osds[i])) { | |
2609 | removed++; | |
2610 | continue; | |
2611 | } | |
2612 | if (removed) { | |
2613 | osds[i - removed] = osds[i]; | |
2614 | } | |
2615 | } | |
2616 | if (removed) | |
2617 | osds.resize(osds.size() - removed); | |
2618 | } else { | |
2619 | for (auto& osd : osds) { | |
2620 | if (!exists(osd)) | |
2621 | osd = CRUSH_ITEM_NONE; | |
2622 | } | |
2623 | } | |
2624 | } | |
2625 | ||
31f18b77 | 2626 | void OSDMap::_pg_to_raw_osds( |
7c673cae FG |
2627 | const pg_pool_t& pool, pg_t pg, |
2628 | vector<int> *osds, | |
2629 | ps_t *ppps) const | |
2630 | { | |
2631 | // map to osds[] | |
2632 | ps_t pps = pool.raw_pg_to_pps(pg); // placement ps | |
2633 | unsigned size = pool.get_size(); | |
2634 | ||
2635 | // what crush rule? | |
20effc67 | 2636 | int ruleno = pool.get_crush_rule(); |
7c673cae FG |
2637 | if (ruleno >= 0) |
2638 | crush->do_rule(ruleno, pps, *osds, size, osd_weight, pg.pool()); | |
2639 | ||
2640 | _remove_nonexistent_osds(pool, *osds); | |
2641 | ||
2642 | if (ppps) | |
2643 | *ppps = pps; | |
7c673cae FG |
2644 | } |
2645 | ||
2646 | int OSDMap::_pick_primary(const vector<int>& osds) const | |
2647 | { | |
2648 | for (auto osd : osds) { | |
2649 | if (osd != CRUSH_ITEM_NONE) { | |
2650 | return osd; | |
2651 | } | |
2652 | } | |
2653 | return -1; | |
2654 | } | |
2655 | ||
224ce89b | 2656 | void OSDMap::_apply_upmap(const pg_pool_t& pi, pg_t raw_pg, vector<int> *raw) const |
7c673cae FG |
2657 | { |
2658 | pg_t pg = pi.raw_pg_to_pg(raw_pg); | |
2659 | auto p = pg_upmap.find(pg); | |
2660 | if (p != pg_upmap.end()) { | |
2661 | // make sure targets aren't marked out | |
2662 | for (auto osd : p->second) { | |
91327a77 AA |
2663 | if (osd != CRUSH_ITEM_NONE && osd < max_osd && osd >= 0 && |
2664 | osd_weight[osd] == 0) { | |
7c673cae FG |
2665 | // reject/ignore the explicit mapping |
2666 | return; | |
2667 | } | |
2668 | } | |
2669 | *raw = vector<int>(p->second.begin(), p->second.end()); | |
224ce89b | 2670 | // continue to check and apply pg_upmap_items if any |
7c673cae FG |
2671 | } |
2672 | ||
2673 | auto q = pg_upmap_items.find(pg); | |
2674 | if (q != pg_upmap_items.end()) { | |
181888fb FG |
2675 | // NOTE: this approach does not allow a bidirectional swap, |
2676 | // e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1]. | |
1e59de90 TL |
2677 | for (auto& [osd_from, osd_to] : q->second) { |
2678 | // A capcaity change upmap (repace osd in the pg with osd not in the pg) | |
181888fb FG |
2679 | // make sure the replacement value doesn't already appear |
2680 | bool exists = false; | |
2681 | ssize_t pos = -1; | |
2682 | for (unsigned i = 0; i < raw->size(); ++i) { | |
2683 | int osd = (*raw)[i]; | |
1e59de90 | 2684 | if (osd == osd_to) { |
181888fb FG |
2685 | exists = true; |
2686 | break; | |
2687 | } | |
2688 | // ignore mapping if target is marked out (or invalid osd id) | |
1e59de90 | 2689 | if (osd == osd_from && |
181888fb | 2690 | pos < 0 && |
1e59de90 TL |
2691 | !(osd_to != CRUSH_ITEM_NONE && osd_to < max_osd && |
2692 | osd_to >= 0 && osd_weight[osd_to] == 0)) { | |
181888fb | 2693 | pos = i; |
1e59de90 | 2694 | } |
181888fb FG |
2695 | } |
2696 | if (!exists && pos >= 0) { | |
1e59de90 TL |
2697 | (*raw)[pos] = osd_to; |
2698 | } | |
2699 | } | |
2700 | } | |
2701 | auto r = pg_upmap_primaries.find(pg); | |
2702 | if (r != pg_upmap_primaries.end()) { | |
2703 | auto new_prim = r->second; | |
2704 | // Apply mapping only if new primary is not marked out and valid osd id | |
2705 | if (new_prim != CRUSH_ITEM_NONE && new_prim < max_osd && new_prim >= 0 && | |
2706 | osd_weight[new_prim] != 0) { | |
2707 | int new_prim_idx = 0; | |
2708 | for (int i = 1 ; i < (int)raw->size(); i++) { // start from 1 on purpose | |
2709 | if ((*raw)[i] == new_prim) { | |
2710 | new_prim_idx = i; | |
2711 | break; | |
2712 | } | |
2713 | } | |
2714 | if (new_prim_idx > 0) { | |
2715 | // swap primary | |
2716 | (*raw)[new_prim_idx] = (*raw)[0]; | |
2717 | (*raw)[0] = new_prim; | |
7c673cae FG |
2718 | } |
2719 | } | |
2720 | } | |
2721 | } | |
2722 | ||
2723 | // pg -> (up osd list) | |
2724 | void OSDMap::_raw_to_up_osds(const pg_pool_t& pool, const vector<int>& raw, | |
2725 | vector<int> *up) const | |
2726 | { | |
2727 | if (pool.can_shift_osds()) { | |
2728 | // shift left | |
2729 | up->clear(); | |
2730 | up->reserve(raw.size()); | |
2731 | for (unsigned i=0; i<raw.size(); i++) { | |
2732 | if (!exists(raw[i]) || is_down(raw[i])) | |
2733 | continue; | |
2734 | up->push_back(raw[i]); | |
2735 | } | |
2736 | } else { | |
2737 | // set down/dne devices to NONE | |
2738 | up->resize(raw.size()); | |
2739 | for (int i = raw.size() - 1; i >= 0; --i) { | |
2740 | if (!exists(raw[i]) || is_down(raw[i])) { | |
2741 | (*up)[i] = CRUSH_ITEM_NONE; | |
2742 | } else { | |
2743 | (*up)[i] = raw[i]; | |
2744 | } | |
2745 | } | |
2746 | } | |
2747 | } | |
2748 | ||
2749 | void OSDMap::_apply_primary_affinity(ps_t seed, | |
2750 | const pg_pool_t& pool, | |
2751 | vector<int> *osds, | |
2752 | int *primary) const | |
2753 | { | |
2754 | // do we have any non-default primary_affinity values for these osds? | |
2755 | if (!osd_primary_affinity) | |
2756 | return; | |
2757 | ||
2758 | bool any = false; | |
2759 | for (const auto osd : *osds) { | |
2760 | if (osd != CRUSH_ITEM_NONE && | |
2761 | (*osd_primary_affinity)[osd] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
2762 | any = true; | |
2763 | break; | |
2764 | } | |
2765 | } | |
2766 | if (!any) | |
2767 | return; | |
2768 | ||
2769 | // pick the primary. feed both the seed (for the pg) and the osd | |
2770 | // into the hash/rng so that a proportional fraction of an osd's pgs | |
2771 | // get rejected as primary. | |
2772 | int pos = -1; | |
2773 | for (unsigned i = 0; i < osds->size(); ++i) { | |
2774 | int o = (*osds)[i]; | |
2775 | if (o == CRUSH_ITEM_NONE) | |
2776 | continue; | |
2777 | unsigned a = (*osd_primary_affinity)[o]; | |
2778 | if (a < CEPH_OSD_MAX_PRIMARY_AFFINITY && | |
2779 | (crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
2780 | seed, o) >> 16) >= a) { | |
2781 | // we chose not to use this primary. note it anyway as a | |
2782 | // fallback in case we don't pick anyone else, but keep looking. | |
2783 | if (pos < 0) | |
2784 | pos = i; | |
2785 | } else { | |
2786 | pos = i; | |
2787 | break; | |
2788 | } | |
2789 | } | |
2790 | if (pos < 0) | |
2791 | return; | |
2792 | ||
2793 | *primary = (*osds)[pos]; | |
2794 | ||
2795 | if (pool.can_shift_osds() && pos > 0) { | |
2796 | // move the new primary to the front. | |
2797 | for (int i = pos; i > 0; --i) { | |
2798 | (*osds)[i] = (*osds)[i-1]; | |
2799 | } | |
2800 | (*osds)[0] = *primary; | |
2801 | } | |
2802 | } | |
2803 | ||
2804 | void OSDMap::_get_temp_osds(const pg_pool_t& pool, pg_t pg, | |
2805 | vector<int> *temp_pg, int *temp_primary) const | |
2806 | { | |
2807 | pg = pool.raw_pg_to_pg(pg); | |
2808 | const auto p = pg_temp->find(pg); | |
2809 | temp_pg->clear(); | |
2810 | if (p != pg_temp->end()) { | |
2811 | for (unsigned i=0; i<p->second.size(); i++) { | |
2812 | if (!exists(p->second[i]) || is_down(p->second[i])) { | |
2813 | if (pool.can_shift_osds()) { | |
2814 | continue; | |
2815 | } else { | |
2816 | temp_pg->push_back(CRUSH_ITEM_NONE); | |
2817 | } | |
2818 | } else { | |
2819 | temp_pg->push_back(p->second[i]); | |
2820 | } | |
2821 | } | |
2822 | } | |
2823 | const auto &pp = primary_temp->find(pg); | |
2824 | *temp_primary = -1; | |
2825 | if (pp != primary_temp->end()) { | |
2826 | *temp_primary = pp->second; | |
2827 | } else if (!temp_pg->empty()) { // apply pg_temp's primary | |
2828 | for (unsigned i = 0; i < temp_pg->size(); ++i) { | |
2829 | if ((*temp_pg)[i] != CRUSH_ITEM_NONE) { | |
2830 | *temp_primary = (*temp_pg)[i]; | |
2831 | break; | |
2832 | } | |
2833 | } | |
2834 | } | |
2835 | } | |
2836 | ||
31f18b77 | 2837 | void OSDMap::pg_to_raw_osds(pg_t pg, vector<int> *raw, int *primary) const |
7c673cae | 2838 | { |
7c673cae | 2839 | const pg_pool_t *pool = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2840 | if (!pool) { |
2841 | *primary = -1; | |
2842 | raw->clear(); | |
31f18b77 | 2843 | return; |
11fdf7f2 | 2844 | } |
31f18b77 | 2845 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
11fdf7f2 | 2846 | *primary = _pick_primary(*raw); |
7c673cae FG |
2847 | } |
2848 | ||
494da23a TL |
2849 | void OSDMap::pg_to_raw_upmap(pg_t pg, vector<int>*raw, |
2850 | vector<int> *raw_upmap) const | |
a8e16298 TL |
2851 | { |
2852 | auto pool = get_pg_pool(pg.pool()); | |
2853 | if (!pool) { | |
2854 | raw_upmap->clear(); | |
2855 | return; | |
2856 | } | |
494da23a TL |
2857 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
2858 | *raw_upmap = *raw; | |
a8e16298 TL |
2859 | _apply_upmap(*pool, pg, raw_upmap); |
2860 | } | |
2861 | ||
7c673cae FG |
2862 | void OSDMap::pg_to_raw_up(pg_t pg, vector<int> *up, int *primary) const |
2863 | { | |
2864 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2865 | if (!pool) { | |
11fdf7f2 TL |
2866 | *primary = -1; |
2867 | up->clear(); | |
7c673cae FG |
2868 | return; |
2869 | } | |
2870 | vector<int> raw; | |
2871 | ps_t pps; | |
2872 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2873 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2874 | _raw_to_up_osds(*pool, raw, up); |
2875 | *primary = _pick_primary(raw); | |
2876 | _apply_primary_affinity(pps, *pool, up, primary); | |
2877 | } | |
31f18b77 | 2878 | |
7c673cae FG |
2879 | void OSDMap::_pg_to_up_acting_osds( |
2880 | const pg_t& pg, vector<int> *up, int *up_primary, | |
2881 | vector<int> *acting, int *acting_primary, | |
2882 | bool raw_pg_to_pg) const | |
2883 | { | |
2884 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2885 | if (!pool || | |
2886 | (!raw_pg_to_pg && pg.ps() >= pool->get_pg_num())) { | |
2887 | if (up) | |
2888 | up->clear(); | |
2889 | if (up_primary) | |
2890 | *up_primary = -1; | |
2891 | if (acting) | |
2892 | acting->clear(); | |
2893 | if (acting_primary) | |
2894 | *acting_primary = -1; | |
2895 | return; | |
2896 | } | |
2897 | vector<int> raw; | |
2898 | vector<int> _up; | |
2899 | vector<int> _acting; | |
2900 | int _up_primary; | |
2901 | int _acting_primary; | |
2902 | ps_t pps; | |
2903 | _get_temp_osds(*pool, pg, &_acting, &_acting_primary); | |
2904 | if (_acting.empty() || up || up_primary) { | |
2905 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2906 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2907 | _raw_to_up_osds(*pool, raw, &_up); |
2908 | _up_primary = _pick_primary(_up); | |
2909 | _apply_primary_affinity(pps, *pool, &_up, &_up_primary); | |
2910 | if (_acting.empty()) { | |
2911 | _acting = _up; | |
2912 | if (_acting_primary == -1) { | |
2913 | _acting_primary = _up_primary; | |
2914 | } | |
2915 | } | |
2916 | ||
2917 | if (up) | |
2918 | up->swap(_up); | |
2919 | if (up_primary) | |
2920 | *up_primary = _up_primary; | |
2921 | } | |
2922 | ||
2923 | if (acting) | |
2924 | acting->swap(_acting); | |
2925 | if (acting_primary) | |
2926 | *acting_primary = _acting_primary; | |
2927 | } | |
2928 | ||
9f95a23c | 2929 | int OSDMap::calc_pg_role_broken(int osd, const vector<int>& acting, int nrep) |
7c673cae | 2930 | { |
9f95a23c TL |
2931 | // This implementation is broken for EC PGs since the osd may appear |
2932 | // multiple times in the acting set. See | |
2933 | // https://tracker.ceph.com/issues/43213 | |
7c673cae FG |
2934 | if (!nrep) |
2935 | nrep = acting.size(); | |
2936 | for (int i=0; i<nrep; i++) | |
2937 | if (acting[i] == osd) | |
2938 | return i; | |
2939 | return -1; | |
2940 | } | |
2941 | ||
9f95a23c | 2942 | int OSDMap::calc_pg_role(pg_shard_t who, const vector<int>& acting) |
7c673cae | 2943 | { |
9f95a23c TL |
2944 | int nrep = acting.size(); |
2945 | if (who.shard == shard_id_t::NO_SHARD) { | |
2946 | for (int i=0; i<nrep; i++) { | |
2947 | if (acting[i] == who.osd) { | |
2948 | return i; | |
2949 | } | |
2950 | } | |
2951 | } else { | |
2952 | if (who.shard < nrep && acting[who.shard] == who.osd) { | |
2953 | return who.shard; | |
2954 | } | |
2955 | } | |
2956 | return -1; | |
7c673cae FG |
2957 | } |
2958 | ||
9f95a23c | 2959 | bool OSDMap::primary_changed_broken( |
7c673cae FG |
2960 | int oldprimary, |
2961 | const vector<int> &oldacting, | |
2962 | int newprimary, | |
2963 | const vector<int> &newacting) | |
2964 | { | |
2965 | if (oldacting.empty() && newacting.empty()) | |
2966 | return false; // both still empty | |
2967 | if (oldacting.empty() ^ newacting.empty()) | |
2968 | return true; // was empty, now not, or vice versa | |
2969 | if (oldprimary != newprimary) | |
2970 | return true; // primary changed | |
9f95a23c TL |
2971 | if (calc_pg_role_broken(oldprimary, oldacting) != |
2972 | calc_pg_role_broken(newprimary, newacting)) | |
7c673cae FG |
2973 | return true; |
2974 | return false; // same primary (tho replicas may have changed) | |
2975 | } | |
2976 | ||
28e407b8 AA |
2977 | uint64_t OSDMap::get_encoding_features() const |
2978 | { | |
2979 | uint64_t f = SIGNIFICANT_FEATURES; | |
df9f7d3d TL |
2980 | if (require_osd_release < ceph_release_t::reef) { |
2981 | f &= ~CEPH_FEATURE_SERVER_REEF; | |
2982 | } | |
9f95a23c TL |
2983 | if (require_osd_release < ceph_release_t::octopus) { |
2984 | f &= ~CEPH_FEATURE_SERVER_OCTOPUS; | |
2985 | } | |
2986 | if (require_osd_release < ceph_release_t::nautilus) { | |
11fdf7f2 TL |
2987 | f &= ~CEPH_FEATURE_SERVER_NAUTILUS; |
2988 | } | |
9f95a23c | 2989 | if (require_osd_release < ceph_release_t::mimic) { |
11fdf7f2 TL |
2990 | f &= ~CEPH_FEATURE_SERVER_MIMIC; |
2991 | } | |
9f95a23c | 2992 | if (require_osd_release < ceph_release_t::luminous) { |
28e407b8 AA |
2993 | f &= ~(CEPH_FEATURE_SERVER_LUMINOUS | |
2994 | CEPH_FEATURE_CRUSH_CHOOSE_ARGS); | |
2995 | } | |
9f95a23c | 2996 | if (require_osd_release < ceph_release_t::kraken) { |
28e407b8 | 2997 | f &= ~(CEPH_FEATURE_SERVER_KRAKEN | |
1adf2230 | 2998 | CEPH_FEATURE_MSG_ADDR2); |
28e407b8 | 2999 | } |
9f95a23c | 3000 | if (require_osd_release < ceph_release_t::jewel) { |
28e407b8 | 3001 | f &= ~(CEPH_FEATURE_SERVER_JEWEL | |
1adf2230 AA |
3002 | CEPH_FEATURE_NEW_OSDOP_ENCODING | |
3003 | CEPH_FEATURE_CRUSH_TUNABLES5); | |
28e407b8 AA |
3004 | } |
3005 | return f; | |
3006 | } | |
7c673cae FG |
3007 | |
3008 | // serialize, unserialize | |
9f95a23c | 3009 | void OSDMap::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 3010 | { |
11fdf7f2 | 3011 | using ceph::encode; |
7c673cae | 3012 | __u16 v = 5; |
11fdf7f2 | 3013 | encode(v, bl); |
7c673cae FG |
3014 | |
3015 | // base | |
11fdf7f2 TL |
3016 | encode(fsid, bl); |
3017 | encode(epoch, bl); | |
3018 | encode(created, bl); | |
3019 | encode(modified, bl); | |
7c673cae | 3020 | |
11fdf7f2 | 3021 | // for encode(pools, bl); |
7c673cae | 3022 | __u32 n = pools.size(); |
11fdf7f2 | 3023 | encode(n, bl); |
7c673cae FG |
3024 | |
3025 | for (const auto &pool : pools) { | |
3026 | n = pool.first; | |
11fdf7f2 TL |
3027 | encode(n, bl); |
3028 | encode(pool.second, bl, 0); | |
7c673cae | 3029 | } |
11fdf7f2 | 3030 | // for encode(pool_name, bl); |
7c673cae | 3031 | n = pool_name.size(); |
11fdf7f2 | 3032 | encode(n, bl); |
7c673cae FG |
3033 | for (const auto &pname : pool_name) { |
3034 | n = pname.first; | |
11fdf7f2 TL |
3035 | encode(n, bl); |
3036 | encode(pname.second, bl); | |
7c673cae | 3037 | } |
11fdf7f2 | 3038 | // for encode(pool_max, bl); |
7c673cae | 3039 | n = pool_max; |
11fdf7f2 | 3040 | encode(n, bl); |
7c673cae | 3041 | |
11fdf7f2 | 3042 | encode(flags, bl); |
7c673cae | 3043 | |
11fdf7f2 | 3044 | encode(max_osd, bl); |
31f18b77 FG |
3045 | { |
3046 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 3047 | encode(n, bl); |
31f18b77 | 3048 | for (auto s : osd_state) { |
11fdf7f2 | 3049 | encode((uint8_t)s, bl); |
31f18b77 FG |
3050 | } |
3051 | } | |
11fdf7f2 TL |
3052 | encode(osd_weight, bl); |
3053 | encode(osd_addrs->client_addrs, bl, 0); | |
7c673cae | 3054 | |
11fdf7f2 | 3055 | // for encode(pg_temp, bl); |
7c673cae | 3056 | n = pg_temp->size(); |
11fdf7f2 | 3057 | encode(n, bl); |
f67539c2 | 3058 | for (const auto& pg : *pg_temp) { |
7c673cae | 3059 | old_pg_t opg = pg.first.get_old_pg(); |
11fdf7f2 TL |
3060 | encode(opg, bl); |
3061 | encode(pg.second, bl); | |
7c673cae FG |
3062 | } |
3063 | ||
3064 | // crush | |
9f95a23c | 3065 | ceph::buffer::list cbl; |
7c673cae | 3066 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 3067 | encode(cbl, bl); |
7c673cae FG |
3068 | } |
3069 | ||
9f95a23c | 3070 | void OSDMap::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 3071 | { |
11fdf7f2 | 3072 | using ceph::encode; |
7c673cae FG |
3073 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
3074 | encode_client_old(bl); | |
3075 | return; | |
3076 | } | |
3077 | ||
3078 | __u16 v = 6; | |
11fdf7f2 | 3079 | encode(v, bl); |
7c673cae FG |
3080 | |
3081 | // base | |
11fdf7f2 TL |
3082 | encode(fsid, bl); |
3083 | encode(epoch, bl); | |
3084 | encode(created, bl); | |
3085 | encode(modified, bl); | |
7c673cae | 3086 | |
11fdf7f2 TL |
3087 | encode(pools, bl, features); |
3088 | encode(pool_name, bl); | |
3089 | encode(pool_max, bl); | |
7c673cae | 3090 | |
11fdf7f2 | 3091 | encode(flags, bl); |
7c673cae | 3092 | |
11fdf7f2 | 3093 | encode(max_osd, bl); |
31f18b77 FG |
3094 | { |
3095 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 3096 | encode(n, bl); |
31f18b77 | 3097 | for (auto s : osd_state) { |
11fdf7f2 | 3098 | encode((uint8_t)s, bl); |
31f18b77 FG |
3099 | } |
3100 | } | |
11fdf7f2 TL |
3101 | encode(osd_weight, bl); |
3102 | encode(osd_addrs->client_addrs, bl, features); | |
7c673cae | 3103 | |
11fdf7f2 | 3104 | encode(*pg_temp, bl); |
7c673cae FG |
3105 | |
3106 | // crush | |
9f95a23c | 3107 | ceph::buffer::list cbl; |
7c673cae | 3108 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 3109 | encode(cbl, bl); |
7c673cae FG |
3110 | |
3111 | // extended | |
3112 | __u16 ev = 10; | |
11fdf7f2 TL |
3113 | encode(ev, bl); |
3114 | encode(osd_addrs->hb_back_addrs, bl, features); | |
3115 | encode(osd_info, bl); | |
f67539c2 | 3116 | encode(blocklist, bl, features); |
11fdf7f2 TL |
3117 | encode(osd_addrs->cluster_addrs, bl, features); |
3118 | encode(cluster_snapshot_epoch, bl); | |
3119 | encode(cluster_snapshot, bl); | |
3120 | encode(*osd_uuid, bl); | |
9f95a23c | 3121 | encode(osd_xinfo, bl, features); |
11fdf7f2 | 3122 | encode(osd_addrs->hb_front_addrs, bl, features); |
7c673cae FG |
3123 | } |
3124 | ||
11fdf7f2 TL |
3125 | /* for a description of osdmap versions, and when they were introduced, please |
3126 | * refer to | |
3127 | * doc/dev/osd_internals/osdmap_versions.txt | |
3128 | */ | |
9f95a23c | 3129 | void OSDMap::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 3130 | { |
11fdf7f2 | 3131 | using ceph::encode; |
7c673cae FG |
3132 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
3133 | encode_classic(bl, features); | |
3134 | return; | |
3135 | } | |
3136 | ||
3137 | // only a select set of callers should *ever* be encoding new | |
3138 | // OSDMaps. others should be passing around the canonical encoded | |
3139 | // buffers from on high. select out those callers by passing in an | |
3140 | // "impossible" feature bit. | |
11fdf7f2 | 3141 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
3142 | features &= ~CEPH_FEATURE_RESERVED; |
3143 | ||
3144 | size_t start_offset = bl.length(); | |
3145 | size_t tail_offset; | |
11fdf7f2 | 3146 | size_t crc_offset; |
9f95a23c | 3147 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
3148 | |
3149 | // meta-encoding: how we include client-used and osd-specific data | |
3150 | ENCODE_START(8, 7, bl); | |
3151 | ||
3152 | { | |
28e407b8 AA |
3153 | // NOTE: any new encoding dependencies must be reflected by |
3154 | // SIGNIFICANT_FEATURES | |
1e59de90 | 3155 | uint8_t v = 10; |
31f18b77 | 3156 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
7c673cae | 3157 | v = 3; |
11fdf7f2 TL |
3158 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
3159 | v = 6; | |
3160 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
3161 | v = 7; | |
df9f7d3d | 3162 | } else if (!HAVE_FEATURE(features, SERVER_REEF)) { |
1e59de90 | 3163 | v = 9; |
df9f7d3d | 3164 | } |
7c673cae FG |
3165 | ENCODE_START(v, 1, bl); // client-usable data |
3166 | // base | |
11fdf7f2 TL |
3167 | encode(fsid, bl); |
3168 | encode(epoch, bl); | |
3169 | encode(created, bl); | |
3170 | encode(modified, bl); | |
7c673cae | 3171 | |
11fdf7f2 TL |
3172 | encode(pools, bl, features); |
3173 | encode(pool_name, bl); | |
3174 | encode(pool_max, bl); | |
7c673cae | 3175 | |
31f18b77 FG |
3176 | if (v < 4) { |
3177 | decltype(flags) f = flags; | |
9f95a23c | 3178 | if (require_osd_release >= ceph_release_t::luminous) |
c07f9fc5 | 3179 | f |= CEPH_OSDMAP_REQUIRE_LUMINOUS | CEPH_OSDMAP_RECOVERY_DELETES; |
9f95a23c | 3180 | else if (require_osd_release == ceph_release_t::kraken) |
31f18b77 | 3181 | f |= CEPH_OSDMAP_REQUIRE_KRAKEN; |
9f95a23c | 3182 | else if (require_osd_release == ceph_release_t::jewel) |
31f18b77 | 3183 | f |= CEPH_OSDMAP_REQUIRE_JEWEL; |
11fdf7f2 | 3184 | encode(f, bl); |
31f18b77 | 3185 | } else { |
11fdf7f2 | 3186 | encode(flags, bl); |
31f18b77 | 3187 | } |
7c673cae | 3188 | |
11fdf7f2 | 3189 | encode(max_osd, bl); |
31f18b77 | 3190 | if (v >= 5) { |
11fdf7f2 | 3191 | encode(osd_state, bl); |
31f18b77 FG |
3192 | } else { |
3193 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 3194 | encode(n, bl); |
31f18b77 | 3195 | for (auto s : osd_state) { |
11fdf7f2 | 3196 | encode((uint8_t)s, bl); |
31f18b77 FG |
3197 | } |
3198 | } | |
11fdf7f2 TL |
3199 | encode(osd_weight, bl); |
3200 | if (v >= 8) { | |
3201 | encode(osd_addrs->client_addrs, bl, features); | |
3202 | } else { | |
3203 | encode_addrvec_pvec_as_addr(osd_addrs->client_addrs, bl, features); | |
3204 | } | |
7c673cae | 3205 | |
11fdf7f2 TL |
3206 | encode(*pg_temp, bl); |
3207 | encode(*primary_temp, bl); | |
7c673cae | 3208 | if (osd_primary_affinity) { |
11fdf7f2 | 3209 | encode(*osd_primary_affinity, bl); |
7c673cae FG |
3210 | } else { |
3211 | vector<__u32> v; | |
11fdf7f2 | 3212 | encode(v, bl); |
7c673cae FG |
3213 | } |
3214 | ||
3215 | // crush | |
9f95a23c | 3216 | ceph::buffer::list cbl; |
7c673cae | 3217 | crush->encode(cbl, features); |
11fdf7f2 TL |
3218 | encode(cbl, bl); |
3219 | encode(erasure_code_profiles, bl); | |
7c673cae FG |
3220 | |
3221 | if (v >= 4) { | |
11fdf7f2 TL |
3222 | encode(pg_upmap, bl); |
3223 | encode(pg_upmap_items, bl); | |
7c673cae | 3224 | } else { |
11fdf7f2 TL |
3225 | ceph_assert(pg_upmap.empty()); |
3226 | ceph_assert(pg_upmap_items.empty()); | |
7c673cae | 3227 | } |
31f18b77 | 3228 | if (v >= 6) { |
11fdf7f2 TL |
3229 | encode(crush_version, bl); |
3230 | } | |
3231 | if (v >= 7) { | |
3232 | encode(new_removed_snaps, bl); | |
3233 | encode(new_purged_snaps, bl); | |
3234 | } | |
3235 | if (v >= 9) { | |
3236 | encode(last_up_change, bl); | |
3237 | encode(last_in_change, bl); | |
31f18b77 | 3238 | } |
1e59de90 TL |
3239 | if (v >= 10) { |
3240 | encode(pg_upmap_primaries, bl); | |
3241 | } else { | |
3242 | ceph_assert(pg_upmap_primaries.empty()); | |
3243 | } | |
7c673cae FG |
3244 | ENCODE_FINISH(bl); // client-usable data |
3245 | } | |
3246 | ||
3247 | { | |
28e407b8 AA |
3248 | // NOTE: any new encoding dependencies must be reflected by |
3249 | // SIGNIFICANT_FEATURES | |
1e59de90 | 3250 | uint8_t target_v = 9; // when bumping this, be aware of allow_crimson |
7c673cae FG |
3251 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
3252 | target_v = 1; | |
11fdf7f2 TL |
3253 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
3254 | target_v = 5; | |
3255 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
3256 | target_v = 6; | |
7c673cae | 3257 | } |
f67539c2 TL |
3258 | if (stretch_mode_enabled) { |
3259 | target_v = std::max((uint8_t)10, target_v); | |
3260 | } | |
33c7a0ef TL |
3261 | if (!range_blocklist.empty()) { |
3262 | target_v = std::max((uint8_t)11, target_v); | |
3263 | } | |
1e59de90 TL |
3264 | if (allow_crimson) { |
3265 | target_v = std::max((uint8_t)12, target_v); | |
3266 | } | |
7c673cae | 3267 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
3268 | if (target_v < 7) { |
3269 | encode_addrvec_pvec_as_addr(osd_addrs->hb_back_addrs, bl, features); | |
3270 | } else { | |
3271 | encode(osd_addrs->hb_back_addrs, bl, features); | |
3272 | } | |
3273 | encode(osd_info, bl); | |
7c673cae FG |
3274 | { |
3275 | // put this in a sorted, ordered map<> so that we encode in a | |
3276 | // deterministic order. | |
f67539c2 TL |
3277 | map<entity_addr_t,utime_t> blocklist_map; |
3278 | for (const auto &addr : blocklist) | |
3279 | blocklist_map.insert(make_pair(addr.first, addr.second)); | |
3280 | encode(blocklist_map, bl, features); | |
11fdf7f2 TL |
3281 | } |
3282 | if (target_v < 7) { | |
3283 | encode_addrvec_pvec_as_addr(osd_addrs->cluster_addrs, bl, features); | |
3284 | } else { | |
3285 | encode(osd_addrs->cluster_addrs, bl, features); | |
3286 | } | |
3287 | encode(cluster_snapshot_epoch, bl); | |
3288 | encode(cluster_snapshot, bl); | |
3289 | encode(*osd_uuid, bl); | |
9f95a23c | 3290 | encode(osd_xinfo, bl, features); |
11fdf7f2 TL |
3291 | if (target_v < 7) { |
3292 | encode_addrvec_pvec_as_addr(osd_addrs->hb_front_addrs, bl, features); | |
3293 | } else { | |
3294 | encode(osd_addrs->hb_front_addrs, bl, features); | |
3295 | } | |
7c673cae | 3296 | if (target_v >= 2) { |
11fdf7f2 TL |
3297 | encode(nearfull_ratio, bl); |
3298 | encode(full_ratio, bl); | |
3299 | encode(backfillfull_ratio, bl); | |
31f18b77 FG |
3300 | } |
3301 | // 4 was string-based new_require_min_compat_client | |
3302 | if (target_v >= 5) { | |
11fdf7f2 TL |
3303 | encode(require_min_compat_client, bl); |
3304 | encode(require_osd_release, bl); | |
3305 | } | |
3306 | if (target_v >= 6) { | |
3307 | encode(removed_snaps_queue, bl); | |
7c673cae | 3308 | } |
81eedcae TL |
3309 | if (target_v >= 8) { |
3310 | encode(crush_node_flags, bl); | |
3311 | } | |
3312 | if (target_v >= 9) { | |
3313 | encode(device_class_flags, bl); | |
3314 | } | |
f67539c2 TL |
3315 | if (target_v >= 10) { |
3316 | encode(stretch_mode_enabled, bl); | |
3317 | encode(stretch_bucket_count, bl); | |
3318 | encode(degraded_stretch_mode, bl); | |
3319 | encode(recovering_stretch_mode, bl); | |
3320 | encode(stretch_mode_bucket, bl); | |
3321 | } | |
33c7a0ef TL |
3322 | if (target_v >= 11) { |
3323 | ::encode(range_blocklist, bl, features); | |
3324 | } | |
1e59de90 TL |
3325 | if (target_v >= 12) { |
3326 | ::encode(allow_crimson, bl); | |
3327 | } | |
7c673cae FG |
3328 | ENCODE_FINISH(bl); // osd-only data |
3329 | } | |
3330 | ||
11fdf7f2 TL |
3331 | crc_offset = bl.length(); |
3332 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
3333 | tail_offset = bl.length(); |
3334 | ||
3335 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
3336 | ||
3337 | // fill in crc | |
9f95a23c | 3338 | ceph::buffer::list front; |
11fdf7f2 | 3339 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae FG |
3340 | crc = front.crc32c(-1); |
3341 | if (tail_offset < bl.length()) { | |
9f95a23c | 3342 | ceph::buffer::list tail; |
7c673cae FG |
3343 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
3344 | crc = tail.crc32c(crc); | |
3345 | } | |
3346 | ceph_le32 crc_le; | |
3347 | crc_le = crc; | |
11fdf7f2 | 3348 | crc_filler->copy_in(4, (char*)&crc_le); |
7c673cae FG |
3349 | crc_defined = true; |
3350 | } | |
3351 | ||
11fdf7f2 TL |
3352 | /* for a description of osdmap versions, and when they were introduced, please |
3353 | * refer to | |
3354 | * doc/dev/osd_internals/osdmap_versions.txt | |
3355 | */ | |
9f95a23c | 3356 | void OSDMap::decode(ceph::buffer::list& bl) |
7c673cae | 3357 | { |
11fdf7f2 | 3358 | auto p = bl.cbegin(); |
7c673cae FG |
3359 | decode(p); |
3360 | } | |
3361 | ||
9f95a23c | 3362 | void OSDMap::decode_classic(ceph::buffer::list::const_iterator& p) |
7c673cae | 3363 | { |
11fdf7f2 | 3364 | using ceph::decode; |
7c673cae FG |
3365 | __u32 n, t; |
3366 | __u16 v; | |
11fdf7f2 | 3367 | decode(v, p); |
7c673cae FG |
3368 | |
3369 | // base | |
11fdf7f2 TL |
3370 | decode(fsid, p); |
3371 | decode(epoch, p); | |
3372 | decode(created, p); | |
3373 | decode(modified, p); | |
7c673cae FG |
3374 | |
3375 | if (v < 6) { | |
3376 | if (v < 4) { | |
3377 | int32_t max_pools = 0; | |
11fdf7f2 | 3378 | decode(max_pools, p); |
7c673cae FG |
3379 | pool_max = max_pools; |
3380 | } | |
3381 | pools.clear(); | |
11fdf7f2 | 3382 | decode(n, p); |
7c673cae | 3383 | while (n--) { |
11fdf7f2 TL |
3384 | decode(t, p); |
3385 | decode(pools[t], p); | |
7c673cae FG |
3386 | } |
3387 | if (v == 4) { | |
11fdf7f2 | 3388 | decode(n, p); |
7c673cae FG |
3389 | pool_max = n; |
3390 | } else if (v == 5) { | |
3391 | pool_name.clear(); | |
11fdf7f2 | 3392 | decode(n, p); |
7c673cae | 3393 | while (n--) { |
11fdf7f2 TL |
3394 | decode(t, p); |
3395 | decode(pool_name[t], p); | |
7c673cae | 3396 | } |
11fdf7f2 | 3397 | decode(n, p); |
7c673cae FG |
3398 | pool_max = n; |
3399 | } | |
3400 | } else { | |
11fdf7f2 TL |
3401 | decode(pools, p); |
3402 | decode(pool_name, p); | |
3403 | decode(pool_max, p); | |
7c673cae FG |
3404 | } |
3405 | // kludge around some old bug that zeroed out pool_max (#2307) | |
3406 | if (pools.size() && pool_max < pools.rbegin()->first) { | |
3407 | pool_max = pools.rbegin()->first; | |
3408 | } | |
3409 | ||
11fdf7f2 | 3410 | decode(flags, p); |
7c673cae | 3411 | |
11fdf7f2 | 3412 | decode(max_osd, p); |
31f18b77 FG |
3413 | { |
3414 | vector<uint8_t> os; | |
11fdf7f2 | 3415 | decode(os, p); |
31f18b77 FG |
3416 | osd_state.resize(os.size()); |
3417 | for (unsigned i = 0; i < os.size(); ++i) { | |
3418 | osd_state[i] = os[i]; | |
3419 | } | |
3420 | } | |
11fdf7f2 TL |
3421 | decode(osd_weight, p); |
3422 | decode(osd_addrs->client_addrs, p); | |
7c673cae FG |
3423 | if (v <= 5) { |
3424 | pg_temp->clear(); | |
11fdf7f2 | 3425 | decode(n, p); |
7c673cae FG |
3426 | while (n--) { |
3427 | old_pg_t opg; | |
9f95a23c | 3428 | ceph::decode_raw(opg, p); |
31f18b77 | 3429 | mempool::osdmap::vector<int32_t> v; |
11fdf7f2 | 3430 | decode(v, p); |
31f18b77 | 3431 | pg_temp->set(pg_t(opg), v); |
7c673cae FG |
3432 | } |
3433 | } else { | |
11fdf7f2 | 3434 | decode(*pg_temp, p); |
7c673cae FG |
3435 | } |
3436 | ||
3437 | // crush | |
9f95a23c | 3438 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3439 | decode(cbl, p); |
3440 | auto cblp = cbl.cbegin(); | |
7c673cae FG |
3441 | crush->decode(cblp); |
3442 | ||
3443 | // extended | |
3444 | __u16 ev = 0; | |
3445 | if (v >= 5) | |
11fdf7f2 TL |
3446 | decode(ev, p); |
3447 | decode(osd_addrs->hb_back_addrs, p); | |
3448 | decode(osd_info, p); | |
7c673cae | 3449 | if (v < 5) |
11fdf7f2 | 3450 | decode(pool_name, p); |
7c673cae | 3451 | |
f67539c2 | 3452 | decode(blocklist, p); |
7c673cae | 3453 | if (ev >= 6) |
11fdf7f2 | 3454 | decode(osd_addrs->cluster_addrs, p); |
7c673cae | 3455 | else |
11fdf7f2 | 3456 | osd_addrs->cluster_addrs.resize(osd_addrs->client_addrs.size()); |
7c673cae FG |
3457 | |
3458 | if (ev >= 7) { | |
11fdf7f2 TL |
3459 | decode(cluster_snapshot_epoch, p); |
3460 | decode(cluster_snapshot, p); | |
7c673cae FG |
3461 | } |
3462 | ||
3463 | if (ev >= 8) { | |
11fdf7f2 | 3464 | decode(*osd_uuid, p); |
7c673cae FG |
3465 | } else { |
3466 | osd_uuid->resize(max_osd); | |
3467 | } | |
3468 | if (ev >= 9) | |
11fdf7f2 | 3469 | decode(osd_xinfo, p); |
7c673cae FG |
3470 | else |
3471 | osd_xinfo.resize(max_osd); | |
3472 | ||
3473 | if (ev >= 10) | |
11fdf7f2 | 3474 | decode(osd_addrs->hb_front_addrs, p); |
7c673cae | 3475 | else |
11fdf7f2 | 3476 | osd_addrs->hb_front_addrs.resize(osd_addrs->hb_back_addrs.size()); |
7c673cae FG |
3477 | |
3478 | osd_primary_affinity.reset(); | |
3479 | ||
3480 | post_decode(); | |
3481 | } | |
3482 | ||
9f95a23c | 3483 | void OSDMap::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 3484 | { |
11fdf7f2 | 3485 | using ceph::decode; |
7c673cae FG |
3486 | /** |
3487 | * Older encodings of the OSDMap had a single struct_v which | |
3488 | * covered the whole encoding, and was prior to our modern | |
3489 | * stuff which includes a compatv and a size. So if we see | |
3490 | * a struct_v < 7, we must rewind to the beginning and use our | |
3491 | * classic decoder. | |
3492 | */ | |
3493 | size_t start_offset = bl.get_off(); | |
3494 | size_t tail_offset = 0; | |
9f95a23c | 3495 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
3496 | |
3497 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
3498 | if (struct_v < 7) { | |
11fdf7f2 | 3499 | bl.seek(start_offset); |
7c673cae FG |
3500 | decode_classic(bl); |
3501 | return; | |
3502 | } | |
3503 | /** | |
3504 | * Since we made it past that hurdle, we can use our normal paths. | |
3505 | */ | |
3506 | { | |
11fdf7f2 | 3507 | DECODE_START(9, bl); // client-usable data |
7c673cae | 3508 | // base |
11fdf7f2 TL |
3509 | decode(fsid, bl); |
3510 | decode(epoch, bl); | |
3511 | decode(created, bl); | |
3512 | decode(modified, bl); | |
7c673cae | 3513 | |
11fdf7f2 TL |
3514 | decode(pools, bl); |
3515 | decode(pool_name, bl); | |
3516 | decode(pool_max, bl); | |
7c673cae | 3517 | |
11fdf7f2 | 3518 | decode(flags, bl); |
7c673cae | 3519 | |
11fdf7f2 | 3520 | decode(max_osd, bl); |
31f18b77 | 3521 | if (struct_v >= 5) { |
11fdf7f2 | 3522 | decode(osd_state, bl); |
31f18b77 FG |
3523 | } else { |
3524 | vector<uint8_t> os; | |
11fdf7f2 | 3525 | decode(os, bl); |
31f18b77 FG |
3526 | osd_state.resize(os.size()); |
3527 | for (unsigned i = 0; i < os.size(); ++i) { | |
3528 | osd_state[i] = os[i]; | |
3529 | } | |
3530 | } | |
11fdf7f2 TL |
3531 | decode(osd_weight, bl); |
3532 | decode(osd_addrs->client_addrs, bl); | |
7c673cae | 3533 | |
11fdf7f2 TL |
3534 | decode(*pg_temp, bl); |
3535 | decode(*primary_temp, bl); | |
3536 | // dates back to firefly. version increased from 2 to 3 still in firefly. | |
3537 | // do we really still need to keep this around? even for old clients? | |
7c673cae FG |
3538 | if (struct_v >= 2) { |
3539 | osd_primary_affinity.reset(new mempool::osdmap::vector<__u32>); | |
11fdf7f2 | 3540 | decode(*osd_primary_affinity, bl); |
7c673cae FG |
3541 | if (osd_primary_affinity->empty()) |
3542 | osd_primary_affinity.reset(); | |
3543 | } else { | |
3544 | osd_primary_affinity.reset(); | |
3545 | } | |
3546 | ||
3547 | // crush | |
9f95a23c | 3548 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3549 | decode(cbl, bl); |
3550 | auto cblp = cbl.cbegin(); | |
7c673cae | 3551 | crush->decode(cblp); |
11fdf7f2 TL |
3552 | // added in firefly; version increased in luminous, so it affects |
3553 | // giant, hammer, infernallis, jewel, and kraken. probably should be left | |
3554 | // alone until we require clients to be all luminous? | |
7c673cae | 3555 | if (struct_v >= 3) { |
11fdf7f2 | 3556 | decode(erasure_code_profiles, bl); |
7c673cae FG |
3557 | } else { |
3558 | erasure_code_profiles.clear(); | |
3559 | } | |
11fdf7f2 TL |
3560 | // version increased from 3 to 4 still in luminous, so same as above |
3561 | // applies. | |
7c673cae | 3562 | if (struct_v >= 4) { |
11fdf7f2 TL |
3563 | decode(pg_upmap, bl); |
3564 | decode(pg_upmap_items, bl); | |
7c673cae FG |
3565 | } else { |
3566 | pg_upmap.clear(); | |
3567 | pg_upmap_items.clear(); | |
3568 | } | |
11fdf7f2 TL |
3569 | // again, version increased from 5 to 6 still in luminous, so above |
3570 | // applies. | |
31f18b77 | 3571 | if (struct_v >= 6) { |
11fdf7f2 TL |
3572 | decode(crush_version, bl); |
3573 | } | |
3574 | // version increase from 6 to 7 in mimic | |
3575 | if (struct_v >= 7) { | |
3576 | decode(new_removed_snaps, bl); | |
3577 | decode(new_purged_snaps, bl); | |
3578 | } | |
3579 | // version increase from 7 to 8, 8 to 9, in nautilus. | |
3580 | if (struct_v >= 9) { | |
3581 | decode(last_up_change, bl); | |
3582 | decode(last_in_change, bl); | |
31f18b77 | 3583 | } |
1e59de90 TL |
3584 | if (struct_v >= 10) { |
3585 | decode(pg_upmap_primaries, bl); | |
3586 | } else { | |
3587 | pg_upmap_primaries.clear(); | |
3588 | } | |
7c673cae FG |
3589 | DECODE_FINISH(bl); // client-usable data |
3590 | } | |
3591 | ||
3592 | { | |
f67539c2 | 3593 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
3594 | decode(osd_addrs->hb_back_addrs, bl); |
3595 | decode(osd_info, bl); | |
f67539c2 | 3596 | decode(blocklist, bl); |
11fdf7f2 TL |
3597 | decode(osd_addrs->cluster_addrs, bl); |
3598 | decode(cluster_snapshot_epoch, bl); | |
3599 | decode(cluster_snapshot, bl); | |
3600 | decode(*osd_uuid, bl); | |
3601 | decode(osd_xinfo, bl); | |
3602 | decode(osd_addrs->hb_front_addrs, bl); | |
3603 | // | |
7c673cae | 3604 | if (struct_v >= 2) { |
11fdf7f2 TL |
3605 | decode(nearfull_ratio, bl); |
3606 | decode(full_ratio, bl); | |
7c673cae FG |
3607 | } else { |
3608 | nearfull_ratio = 0; | |
3609 | full_ratio = 0; | |
3610 | } | |
3611 | if (struct_v >= 3) { | |
11fdf7f2 | 3612 | decode(backfillfull_ratio, bl); |
7c673cae FG |
3613 | } else { |
3614 | backfillfull_ratio = 0; | |
3615 | } | |
31f18b77 FG |
3616 | if (struct_v == 4) { |
3617 | string r; | |
11fdf7f2 | 3618 | decode(r, bl); |
31f18b77 FG |
3619 | if (r.length()) |
3620 | require_min_compat_client = ceph_release_from_name(r.c_str()); | |
3621 | } | |
3622 | if (struct_v >= 5) { | |
11fdf7f2 TL |
3623 | decode(require_min_compat_client, bl); |
3624 | decode(require_osd_release, bl); | |
9f95a23c | 3625 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
3626 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
3627 | } | |
9f95a23c | 3628 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 3629 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3630 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
3631 | } |
3632 | } else { | |
3633 | if (flags & CEPH_OSDMAP_REQUIRE_LUMINOUS) { | |
3634 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 3635 | require_osd_release = ceph_release_t::luminous; |
31f18b77 | 3636 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3637 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 | 3638 | } else if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { |
9f95a23c | 3639 | require_osd_release = ceph_release_t::kraken; |
31f18b77 | 3640 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { |
9f95a23c | 3641 | require_osd_release = ceph_release_t::jewel; |
31f18b77 | 3642 | } else { |
9f95a23c | 3643 | require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
3644 | } |
3645 | } | |
11fdf7f2 TL |
3646 | if (struct_v >= 6) { |
3647 | decode(removed_snaps_queue, bl); | |
3648 | } | |
81eedcae TL |
3649 | if (struct_v >= 8) { |
3650 | decode(crush_node_flags, bl); | |
3651 | } else { | |
3652 | crush_node_flags.clear(); | |
3653 | } | |
3654 | if (struct_v >= 9) { | |
3655 | decode(device_class_flags, bl); | |
3656 | } else { | |
3657 | device_class_flags.clear(); | |
3658 | } | |
f67539c2 TL |
3659 | if (struct_v >= 10) { |
3660 | decode(stretch_mode_enabled, bl); | |
3661 | decode(stretch_bucket_count, bl); | |
3662 | decode(degraded_stretch_mode, bl); | |
3663 | decode(recovering_stretch_mode, bl); | |
3664 | decode(stretch_mode_bucket, bl); | |
3665 | } else { | |
3666 | stretch_mode_enabled = false; | |
3667 | stretch_bucket_count = 0; | |
3668 | degraded_stretch_mode = 0; | |
3669 | recovering_stretch_mode = 0; | |
3670 | stretch_mode_bucket = 0; | |
3671 | } | |
33c7a0ef TL |
3672 | if (struct_v >= 11) { |
3673 | decode(range_blocklist, bl); | |
3674 | calculated_ranges.clear(); | |
3675 | for (const auto& i : range_blocklist) { | |
3676 | calculated_ranges.emplace(i.first, i.first); | |
3677 | } | |
3678 | } | |
1e59de90 TL |
3679 | if (struct_v >= 12) { |
3680 | decode(allow_crimson, bl); | |
3681 | } | |
7c673cae FG |
3682 | DECODE_FINISH(bl); // osd-only data |
3683 | } | |
3684 | ||
3685 | if (struct_v >= 8) { | |
3686 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 3687 | decode(crc, bl); |
7c673cae FG |
3688 | tail_offset = bl.get_off(); |
3689 | crc_defined = true; | |
3690 | } else { | |
3691 | crc_defined = false; | |
3692 | crc = 0; | |
3693 | } | |
3694 | ||
3695 | DECODE_FINISH(bl); // wrapper | |
3696 | ||
3697 | if (tail_offset) { | |
3698 | // verify crc | |
3699 | uint32_t actual = crc_front.crc32c(-1); | |
3700 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 3701 | ceph::buffer::list tail; |
7c673cae FG |
3702 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
3703 | actual = tail.crc32c(actual); | |
3704 | } | |
3705 | if (crc != actual) { | |
3706 | ostringstream ss; | |
3707 | ss << "bad crc, actual " << actual << " != expected " << crc; | |
3708 | string s = ss.str(); | |
9f95a23c | 3709 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
3710 | } |
3711 | } | |
3712 | ||
3713 | post_decode(); | |
3714 | } | |
3715 | ||
3716 | void OSDMap::post_decode() | |
3717 | { | |
3718 | // index pool names | |
3719 | name_pool.clear(); | |
3720 | for (const auto &pname : pool_name) { | |
3721 | name_pool[pname.second] = pname.first; | |
3722 | } | |
3723 | ||
3724 | calc_num_osds(); | |
3725 | _calc_up_osd_features(); | |
3726 | } | |
3727 | ||
3728 | void OSDMap::dump_erasure_code_profiles( | |
3729 | const mempool::osdmap::map<string,map<string,string>>& profiles, | |
3730 | Formatter *f) | |
3731 | { | |
3732 | f->open_object_section("erasure_code_profiles"); | |
3733 | for (const auto &profile : profiles) { | |
3734 | f->open_object_section(profile.first.c_str()); | |
3735 | for (const auto &profm : profile.second) { | |
9f95a23c | 3736 | f->dump_string(profm.first.c_str(), profm.second); |
7c673cae FG |
3737 | } |
3738 | f->close_section(); | |
3739 | } | |
3740 | f->close_section(); | |
3741 | } | |
3742 | ||
9f95a23c TL |
3743 | void OSDMap::dump_osds(Formatter *f) const |
3744 | { | |
3745 | f->open_array_section("osds"); | |
3746 | for (int i=0; i<get_max_osd(); i++) { | |
3747 | if (exists(i)) { | |
3748 | dump_osd(i, f); | |
3749 | } | |
3750 | } | |
3751 | f->close_section(); | |
3752 | } | |
3753 | ||
3754 | void OSDMap::dump_osd(int id, Formatter *f) const | |
3755 | { | |
3756 | ceph_assert(f != nullptr); | |
3757 | if (!exists(id)) { | |
3758 | return; | |
3759 | } | |
3760 | ||
3761 | f->open_object_section("osd_info"); | |
3762 | f->dump_int("osd", id); | |
3763 | f->dump_stream("uuid") << get_uuid(id); | |
3764 | f->dump_int("up", is_up(id)); | |
3765 | f->dump_int("in", is_in(id)); | |
3766 | f->dump_float("weight", get_weightf(id)); | |
3767 | f->dump_float("primary_affinity", get_primary_affinityf(id)); | |
3768 | get_info(id).dump(f); | |
3769 | f->dump_object("public_addrs", get_addrs(id)); | |
3770 | f->dump_object("cluster_addrs", get_cluster_addrs(id)); | |
3771 | f->dump_object("heartbeat_back_addrs", get_hb_back_addrs(id)); | |
3772 | f->dump_object("heartbeat_front_addrs", get_hb_front_addrs(id)); | |
3773 | // compat | |
3774 | f->dump_stream("public_addr") << get_addrs(id).get_legacy_str(); | |
3775 | f->dump_stream("cluster_addr") << get_cluster_addrs(id).get_legacy_str(); | |
3776 | f->dump_stream("heartbeat_back_addr") | |
3777 | << get_hb_back_addrs(id).get_legacy_str(); | |
3778 | f->dump_stream("heartbeat_front_addr") | |
3779 | << get_hb_front_addrs(id).get_legacy_str(); | |
3780 | ||
3781 | set<string> st; | |
3782 | get_state(id, st); | |
3783 | f->open_array_section("state"); | |
3784 | for (const auto &state : st) | |
3785 | f->dump_string("state", state); | |
3786 | f->close_section(); | |
3787 | ||
3788 | f->close_section(); | |
3789 | } | |
3790 | ||
1e59de90 TL |
3791 | void OSDMap::dump_pool(CephContext *cct, |
3792 | int64_t pid, | |
3793 | const pg_pool_t &pdata, | |
3794 | ceph::Formatter *f) const | |
3795 | { | |
3796 | std::string name("<unknown>"); | |
3797 | const auto &pni = pool_name.find(pid); | |
3798 | if (pni != pool_name.end()) | |
3799 | name = pni->second; | |
3800 | f->open_object_section("pool"); | |
3801 | f->dump_int("pool", pid); | |
3802 | f->dump_string("pool_name", name); | |
3803 | pdata.dump(f); | |
3804 | dump_read_balance_score(cct, pid, pdata, f); | |
3805 | f->close_section(); // pool | |
3806 | } | |
3807 | ||
3808 | void OSDMap::dump_read_balance_score(CephContext *cct, | |
3809 | int64_t pid, | |
3810 | const pg_pool_t &pdata, | |
3811 | ceph::Formatter *f) const | |
3812 | { | |
3813 | if (pdata.is_replicated()) { | |
3814 | // Add rb section with values for score, optimal score, raw score | |
3815 | // // and primary_affinity average | |
3816 | OSDMap::read_balance_info_t rb_info; | |
3817 | auto rc = calc_read_balance_score(cct, pid, &rb_info); | |
3818 | if (rc >= 0) { | |
3819 | f->open_object_section("read_balance"); | |
3820 | f->dump_float("score_acting", rb_info.acting_adj_score); | |
3821 | f->dump_float("score_stable", rb_info.adjusted_score); | |
3822 | f->dump_float("optimal_score", rb_info.optimal_score); | |
3823 | f->dump_float("raw_score_acting", rb_info.acting_raw_score); | |
3824 | f->dump_float("raw_score_stable", rb_info.raw_score); | |
3825 | f->dump_float("primary_affinity_weighted", rb_info.pa_weighted); | |
3826 | f->dump_float("average_primary_affinity", rb_info.pa_avg); | |
3827 | f->dump_float("average_primary_affinity_weighted", rb_info.pa_weighted_avg); | |
3828 | if (rb_info.err_msg.length() > 0) { | |
3829 | f->dump_string("error_message", rb_info.err_msg); | |
3830 | } | |
3831 | f->close_section(); // read_balance | |
3832 | } | |
3833 | else { | |
3834 | if (rb_info.err_msg.length() > 0) { | |
3835 | f->open_object_section("read_balance"); | |
3836 | f->dump_string("error_message", rb_info.err_msg); | |
3837 | f->dump_float("score_acting", rb_info.acting_adj_score); | |
3838 | f->dump_float("score_stable", rb_info.adjusted_score); | |
3839 | f->close_section(); // read_balance | |
3840 | } | |
3841 | } | |
3842 | } | |
3843 | } | |
3844 | ||
3845 | void OSDMap::dump(Formatter *f, CephContext *cct) const | |
7c673cae FG |
3846 | { |
3847 | f->dump_int("epoch", get_epoch()); | |
3848 | f->dump_stream("fsid") << get_fsid(); | |
3849 | f->dump_stream("created") << get_created(); | |
3850 | f->dump_stream("modified") << get_modified(); | |
11fdf7f2 TL |
3851 | f->dump_stream("last_up_change") << last_up_change; |
3852 | f->dump_stream("last_in_change") << last_in_change; | |
7c673cae | 3853 | f->dump_string("flags", get_flag_string()); |
11fdf7f2 TL |
3854 | f->dump_unsigned("flags_num", flags); |
3855 | f->open_array_section("flags_set"); | |
3856 | set<string> flagset; | |
3857 | get_flag_set(&flagset); | |
3858 | for (auto p : flagset) { | |
3859 | f->dump_string("flag", p); | |
3860 | } | |
3861 | f->close_section(); | |
31f18b77 | 3862 | f->dump_unsigned("crush_version", get_crush_version()); |
7c673cae FG |
3863 | f->dump_float("full_ratio", full_ratio); |
3864 | f->dump_float("backfillfull_ratio", backfillfull_ratio); | |
3865 | f->dump_float("nearfull_ratio", nearfull_ratio); | |
3866 | f->dump_string("cluster_snapshot", get_cluster_snapshot()); | |
3867 | f->dump_int("pool_max", get_pool_max()); | |
3868 | f->dump_int("max_osd", get_max_osd()); | |
31f18b77 | 3869 | f->dump_string("require_min_compat_client", |
f67539c2 | 3870 | to_string(require_min_compat_client)); |
31f18b77 | 3871 | f->dump_string("min_compat_client", |
f67539c2 | 3872 | to_string(get_min_compat_client())); |
31f18b77 | 3873 | f->dump_string("require_osd_release", |
f67539c2 | 3874 | to_string(require_osd_release)); |
7c673cae | 3875 | |
1e59de90 | 3876 | f->dump_bool("allow_crimson", allow_crimson); |
7c673cae | 3877 | f->open_array_section("pools"); |
1e59de90 TL |
3878 | for (const auto &[pid, pdata] : pools) { |
3879 | dump_pool(cct, pid, pdata, f); | |
7c673cae FG |
3880 | } |
3881 | f->close_section(); | |
3882 | ||
9f95a23c | 3883 | dump_osds(f); |
7c673cae FG |
3884 | |
3885 | f->open_array_section("osd_xinfo"); | |
3886 | for (int i=0; i<get_max_osd(); i++) { | |
3887 | if (exists(i)) { | |
3888 | f->open_object_section("xinfo"); | |
3889 | f->dump_int("osd", i); | |
3890 | osd_xinfo[i].dump(f); | |
3891 | f->close_section(); | |
3892 | } | |
3893 | } | |
3894 | f->close_section(); | |
3895 | ||
3896 | f->open_array_section("pg_upmap"); | |
3897 | for (auto& p : pg_upmap) { | |
3898 | f->open_object_section("mapping"); | |
3899 | f->dump_stream("pgid") << p.first; | |
3900 | f->open_array_section("osds"); | |
3901 | for (auto q : p.second) { | |
3902 | f->dump_int("osd", q); | |
3903 | } | |
3904 | f->close_section(); | |
3905 | f->close_section(); | |
3906 | } | |
3907 | f->close_section(); | |
1e59de90 | 3908 | |
7c673cae | 3909 | f->open_array_section("pg_upmap_items"); |
1e59de90 | 3910 | for (auto& [pgid, mappings] : pg_upmap_items) { |
7c673cae | 3911 | f->open_object_section("mapping"); |
1e59de90 | 3912 | f->dump_stream("pgid") << pgid; |
7c673cae | 3913 | f->open_array_section("mappings"); |
1e59de90 | 3914 | for (auto& [from, to] : mappings) { |
7c673cae | 3915 | f->open_object_section("mapping"); |
1e59de90 TL |
3916 | f->dump_int("from", from); |
3917 | f->dump_int("to", to); | |
7c673cae FG |
3918 | f->close_section(); |
3919 | } | |
3920 | f->close_section(); | |
3921 | f->close_section(); | |
3922 | } | |
3923 | f->close_section(); | |
1e59de90 TL |
3924 | |
3925 | f->open_array_section("pg_upmap_primaries"); | |
3926 | for (const auto& [pg, osd] : pg_upmap_primaries) { | |
3927 | f->open_object_section("primary_mapping"); | |
3928 | f->dump_stream("pgid") << pg; | |
3929 | f->dump_int("primary_osd", osd); | |
3930 | f->close_section(); | |
3931 | } | |
3932 | f->close_section(); // primary_temp | |
3933 | ||
7c673cae | 3934 | f->open_array_section("pg_temp"); |
31f18b77 | 3935 | pg_temp->dump(f); |
7c673cae FG |
3936 | f->close_section(); |
3937 | ||
3938 | f->open_array_section("primary_temp"); | |
3939 | for (const auto &pg : *primary_temp) { | |
3940 | f->dump_stream("pgid") << pg.first; | |
3941 | f->dump_int("osd", pg.second); | |
3942 | } | |
3943 | f->close_section(); // primary_temp | |
3944 | ||
f67539c2 TL |
3945 | f->open_object_section("blocklist"); |
3946 | for (const auto &addr : blocklist) { | |
7c673cae FG |
3947 | stringstream ss; |
3948 | ss << addr.first; | |
3949 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3950 | } | |
3951 | f->close_section(); | |
33c7a0ef TL |
3952 | f->open_object_section("range_blocklist"); |
3953 | for (const auto &addr : range_blocklist) { | |
3954 | stringstream ss; | |
3955 | ss << addr.first; | |
3956 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3957 | } | |
3958 | f->close_section(); | |
7c673cae FG |
3959 | |
3960 | dump_erasure_code_profiles(erasure_code_profiles, f); | |
11fdf7f2 TL |
3961 | |
3962 | f->open_array_section("removed_snaps_queue"); | |
3963 | for (auto& p : removed_snaps_queue) { | |
3964 | f->open_object_section("pool"); | |
3965 | f->dump_int("pool", p.first); | |
3966 | f->open_array_section("snaps"); | |
3967 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3968 | f->open_object_section("interval"); | |
3969 | f->dump_unsigned("begin", q.get_start()); | |
3970 | f->dump_unsigned("length", q.get_len()); | |
3971 | f->close_section(); | |
3972 | } | |
3973 | f->close_section(); | |
3974 | f->close_section(); | |
3975 | } | |
3976 | f->close_section(); | |
3977 | f->open_array_section("new_removed_snaps"); | |
3978 | for (auto& p : new_removed_snaps) { | |
3979 | f->open_object_section("pool"); | |
3980 | f->dump_int("pool", p.first); | |
3981 | f->open_array_section("snaps"); | |
3982 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3983 | f->open_object_section("interval"); | |
3984 | f->dump_unsigned("begin", q.get_start()); | |
3985 | f->dump_unsigned("length", q.get_len()); | |
3986 | f->close_section(); | |
3987 | } | |
3988 | f->close_section(); | |
3989 | f->close_section(); | |
3990 | } | |
3991 | f->close_section(); | |
3992 | f->open_array_section("new_purged_snaps"); | |
3993 | for (auto& p : new_purged_snaps) { | |
3994 | f->open_object_section("pool"); | |
3995 | f->dump_int("pool", p.first); | |
3996 | f->open_array_section("snaps"); | |
3997 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3998 | f->open_object_section("interval"); | |
3999 | f->dump_unsigned("begin", q.get_start()); | |
4000 | f->dump_unsigned("length", q.get_len()); | |
4001 | f->close_section(); | |
4002 | } | |
4003 | f->close_section(); | |
4004 | f->close_section(); | |
4005 | } | |
4006 | f->close_section(); | |
81eedcae TL |
4007 | f->open_object_section("crush_node_flags"); |
4008 | for (auto& i : crush_node_flags) { | |
4009 | string s = crush->item_exists(i.first) ? crush->get_item_name(i.first) | |
4010 | : stringify(i.first); | |
4011 | f->open_array_section(s.c_str()); | |
4012 | set<string> st; | |
4013 | calc_state_set(i.second, st); | |
4014 | for (auto& j : st) { | |
4015 | f->dump_string("flag", j); | |
4016 | } | |
4017 | f->close_section(); | |
4018 | } | |
4019 | f->close_section(); | |
4020 | f->open_object_section("device_class_flags"); | |
4021 | for (auto& i : device_class_flags) { | |
4022 | const char* class_name = crush->get_class_name(i.first); | |
4023 | string s = class_name ? class_name : stringify(i.first); | |
4024 | f->open_array_section(s.c_str()); | |
4025 | set<string> st; | |
4026 | calc_state_set(i.second, st); | |
4027 | for (auto& j : st) { | |
4028 | f->dump_string("flag", j); | |
4029 | } | |
4030 | f->close_section(); | |
4031 | } | |
4032 | f->close_section(); | |
f67539c2 TL |
4033 | f->open_object_section("stretch_mode"); |
4034 | { | |
4035 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
4036 | f->dump_unsigned("stretch_bucket_count", stretch_bucket_count); | |
4037 | f->dump_unsigned("degraded_stretch_mode", degraded_stretch_mode); | |
4038 | f->dump_unsigned("recovering_stretch_mode", recovering_stretch_mode); | |
4039 | f->dump_int("stretch_mode_bucket", stretch_mode_bucket); | |
4040 | } | |
4041 | f->close_section(); | |
7c673cae FG |
4042 | } |
4043 | ||
4044 | void OSDMap::generate_test_instances(list<OSDMap*>& o) | |
4045 | { | |
4046 | o.push_back(new OSDMap); | |
4047 | ||
4048 | CephContext *cct = new CephContext(CODE_ENVIRONMENT_UTILITY); | |
4049 | o.push_back(new OSDMap); | |
4050 | uuid_d fsid; | |
224ce89b | 4051 | o.back()->build_simple(cct, 1, fsid, 16); |
7c673cae | 4052 | o.back()->created = o.back()->modified = utime_t(1, 2); // fix timestamp |
f67539c2 | 4053 | o.back()->blocklist[entity_addr_t()] = utime_t(5, 6); |
7c673cae FG |
4054 | cct->put(); |
4055 | } | |
4056 | ||
4057 | string OSDMap::get_flag_string(unsigned f) | |
4058 | { | |
4059 | string s; | |
7c673cae FG |
4060 | if (f & CEPH_OSDMAP_PAUSERD) |
4061 | s += ",pauserd"; | |
4062 | if (f & CEPH_OSDMAP_PAUSEWR) | |
4063 | s += ",pausewr"; | |
4064 | if (f & CEPH_OSDMAP_PAUSEREC) | |
4065 | s += ",pauserec"; | |
4066 | if (f & CEPH_OSDMAP_NOUP) | |
4067 | s += ",noup"; | |
4068 | if (f & CEPH_OSDMAP_NODOWN) | |
4069 | s += ",nodown"; | |
4070 | if (f & CEPH_OSDMAP_NOOUT) | |
4071 | s += ",noout"; | |
4072 | if (f & CEPH_OSDMAP_NOIN) | |
4073 | s += ",noin"; | |
4074 | if (f & CEPH_OSDMAP_NOBACKFILL) | |
4075 | s += ",nobackfill"; | |
4076 | if (f & CEPH_OSDMAP_NOREBALANCE) | |
4077 | s += ",norebalance"; | |
4078 | if (f & CEPH_OSDMAP_NORECOVER) | |
4079 | s += ",norecover"; | |
4080 | if (f & CEPH_OSDMAP_NOSCRUB) | |
4081 | s += ",noscrub"; | |
4082 | if (f & CEPH_OSDMAP_NODEEP_SCRUB) | |
4083 | s += ",nodeep-scrub"; | |
4084 | if (f & CEPH_OSDMAP_NOTIERAGENT) | |
4085 | s += ",notieragent"; | |
11fdf7f2 TL |
4086 | if (f & CEPH_OSDMAP_NOSNAPTRIM) |
4087 | s += ",nosnaptrim"; | |
7c673cae FG |
4088 | if (f & CEPH_OSDMAP_SORTBITWISE) |
4089 | s += ",sortbitwise"; | |
4090 | if (f & CEPH_OSDMAP_REQUIRE_JEWEL) | |
4091 | s += ",require_jewel_osds"; | |
4092 | if (f & CEPH_OSDMAP_REQUIRE_KRAKEN) | |
4093 | s += ",require_kraken_osds"; | |
4094 | if (f & CEPH_OSDMAP_REQUIRE_LUMINOUS) | |
4095 | s += ",require_luminous_osds"; | |
c07f9fc5 FG |
4096 | if (f & CEPH_OSDMAP_RECOVERY_DELETES) |
4097 | s += ",recovery_deletes"; | |
181888fb FG |
4098 | if (f & CEPH_OSDMAP_PURGED_SNAPDIRS) |
4099 | s += ",purged_snapdirs"; | |
f64942e4 AA |
4100 | if (f & CEPH_OSDMAP_PGLOG_HARDLIMIT) |
4101 | s += ",pglog_hardlimit"; | |
aee94f69 TL |
4102 | if (f & CEPH_OSDMAP_NOAUTOSCALE) |
4103 | s += ",noautoscale"; | |
7c673cae FG |
4104 | if (s.length()) |
4105 | s.erase(0, 1); | |
4106 | return s; | |
4107 | } | |
4108 | ||
4109 | string OSDMap::get_flag_string() const | |
4110 | { | |
4111 | return get_flag_string(flags); | |
4112 | } | |
4113 | ||
1e59de90 | 4114 | void OSDMap::print_pools(CephContext *cct, ostream& out) const |
7c673cae | 4115 | { |
1e59de90 | 4116 | for (const auto &[pid, pdata] : pools) { |
7c673cae | 4117 | std::string name("<unknown>"); |
1e59de90 | 4118 | const auto &pni = pool_name.find(pid); |
7c673cae FG |
4119 | if (pni != pool_name.end()) |
4120 | name = pni->second; | |
1e59de90 TL |
4121 | char rb_score_str[32] = ""; |
4122 | int rc = 0; | |
4123 | read_balance_info_t rb_info; | |
4124 | if (pdata.is_replicated()) { | |
4125 | rc = calc_read_balance_score(cct, pid, &rb_info); | |
4126 | if (rc >= 0) | |
4127 | snprintf (rb_score_str, sizeof(rb_score_str), | |
4128 | " read_balance_score %.2f", rb_info.acting_adj_score); | |
4129 | } | |
4130 | ||
4131 | out << "pool " << pid | |
7c673cae | 4132 | << " '" << name |
1e59de90 TL |
4133 | << "' " << pdata |
4134 | << rb_score_str << "\n"; | |
4135 | if (rb_info.err_msg.length() > 0) { | |
4136 | out << (rc < 0 ? " ERROR: " : " Warning: ") << rb_info.err_msg << "\n"; | |
4137 | } | |
4138 | ||
4139 | //TODO - print error messages here. | |
7c673cae | 4140 | |
1e59de90 | 4141 | for (const auto &snap : pdata.snaps) |
7c673cae FG |
4142 | out << "\tsnap " << snap.second.snapid << " '" << snap.second.name << "' " << snap.second.stamp << "\n"; |
4143 | ||
1e59de90 TL |
4144 | if (!pdata.removed_snaps.empty()) |
4145 | out << "\tremoved_snaps " << pdata.removed_snaps << "\n"; | |
4146 | auto p = removed_snaps_queue.find(pid); | |
11fdf7f2 TL |
4147 | if (p != removed_snaps_queue.end()) { |
4148 | out << "\tremoved_snaps_queue " << p->second << "\n"; | |
4149 | } | |
7c673cae FG |
4150 | } |
4151 | out << std::endl; | |
4152 | } | |
4153 | ||
9f95a23c TL |
4154 | void OSDMap::print_osds(ostream& out) const |
4155 | { | |
4156 | for (int i=0; i<get_max_osd(); i++) { | |
4157 | if (exists(i)) { | |
4158 | print_osd(i, out); | |
4159 | } | |
4160 | } | |
4161 | } | |
4162 | void OSDMap::print_osd(int id, ostream& out) const | |
4163 | { | |
4164 | if (!exists(id)) { | |
4165 | return; | |
4166 | } | |
4167 | ||
4168 | out << "osd." << id; | |
4169 | out << (is_up(id) ? " up ":" down"); | |
4170 | out << (is_in(id) ? " in ":" out"); | |
4171 | out << " weight " << get_weightf(id); | |
4172 | if (get_primary_affinity(id) != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
4173 | out << " primary_affinity " << get_primary_affinityf(id); | |
4174 | } | |
4175 | const osd_info_t& info(get_info(id)); | |
4176 | out << " " << info; | |
4177 | out << " " << get_addrs(id) << " " << get_cluster_addrs(id); | |
4178 | set<string> st; | |
4179 | get_state(id, st); | |
4180 | out << " " << st; | |
4181 | if (!get_uuid(id).is_zero()) { | |
4182 | out << " " << get_uuid(id); | |
4183 | } | |
4184 | out << "\n"; | |
4185 | } | |
4186 | ||
1e59de90 | 4187 | void OSDMap::print(CephContext *cct, ostream& out) const |
7c673cae FG |
4188 | { |
4189 | out << "epoch " << get_epoch() << "\n" | |
4190 | << "fsid " << get_fsid() << "\n" | |
4191 | << "created " << get_created() << "\n" | |
4192 | << "modified " << get_modified() << "\n"; | |
4193 | ||
4194 | out << "flags " << get_flag_string() << "\n"; | |
31f18b77 | 4195 | out << "crush_version " << get_crush_version() << "\n"; |
7c673cae FG |
4196 | out << "full_ratio " << full_ratio << "\n"; |
4197 | out << "backfillfull_ratio " << backfillfull_ratio << "\n"; | |
4198 | out << "nearfull_ratio " << nearfull_ratio << "\n"; | |
9f95a23c | 4199 | if (require_min_compat_client != ceph_release_t::unknown) { |
31f18b77 | 4200 | out << "require_min_compat_client " |
9f95a23c | 4201 | << require_min_compat_client << "\n"; |
7c673cae | 4202 | } |
9f95a23c | 4203 | out << "min_compat_client " << get_min_compat_client() |
31f18b77 | 4204 | << "\n"; |
9f95a23c TL |
4205 | if (require_osd_release > ceph_release_t::unknown) { |
4206 | out << "require_osd_release " << require_osd_release | |
224ce89b WB |
4207 | << "\n"; |
4208 | } | |
f67539c2 TL |
4209 | out << "stretch_mode_enabled " << (stretch_mode_enabled ? "true" : "false") << "\n"; |
4210 | if (stretch_mode_enabled) { | |
4211 | out << "stretch_bucket_count " << stretch_bucket_count << "\n"; | |
4212 | out << "degraded_stretch_mode " << degraded_stretch_mode << "\n"; | |
4213 | out << "recovering_stretch_mode " << recovering_stretch_mode << "\n"; | |
4214 | out << "stretch_mode_bucket " << stretch_mode_bucket << "\n"; | |
4215 | } | |
7c673cae FG |
4216 | if (get_cluster_snapshot().length()) |
4217 | out << "cluster_snapshot " << get_cluster_snapshot() << "\n"; | |
1e59de90 TL |
4218 | if (allow_crimson) { |
4219 | out << "allow_crimson=true\n"; | |
4220 | } | |
7c673cae FG |
4221 | out << "\n"; |
4222 | ||
1e59de90 | 4223 | print_pools(cct, out); |
7c673cae FG |
4224 | |
4225 | out << "max_osd " << get_max_osd() << "\n"; | |
9f95a23c | 4226 | print_osds(out); |
7c673cae FG |
4227 | out << std::endl; |
4228 | ||
4229 | for (auto& p : pg_upmap) { | |
4230 | out << "pg_upmap " << p.first << " " << p.second << "\n"; | |
4231 | } | |
4232 | for (auto& p : pg_upmap_items) { | |
4233 | out << "pg_upmap_items " << p.first << " " << p.second << "\n"; | |
4234 | } | |
4235 | ||
1e59de90 TL |
4236 | for (auto& [pg, osd] : pg_upmap_primaries) { |
4237 | out << "pg_upmap_primary " << pg << " " << osd << "\n"; | |
4238 | } | |
4239 | ||
f67539c2 | 4240 | for (const auto& pg : *pg_temp) |
7c673cae FG |
4241 | out << "pg_temp " << pg.first << " " << pg.second << "\n"; |
4242 | ||
f67539c2 | 4243 | for (const auto& pg : *primary_temp) |
7c673cae FG |
4244 | out << "primary_temp " << pg.first << " " << pg.second << "\n"; |
4245 | ||
f67539c2 TL |
4246 | for (const auto &addr : blocklist) |
4247 | out << "blocklist " << addr.first << " expires " << addr.second << "\n"; | |
33c7a0ef TL |
4248 | for (const auto &addr : range_blocklist) |
4249 | out << "range blocklist " << addr.first << " expires " << addr.second << "\n"; | |
7c673cae FG |
4250 | } |
4251 | ||
4252 | class OSDTreePlainDumper : public CrushTreeDumper::Dumper<TextTable> { | |
4253 | public: | |
4254 | typedef CrushTreeDumper::Dumper<TextTable> Parent; | |
31f18b77 FG |
4255 | |
4256 | OSDTreePlainDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
4257 | unsigned f) | |
c07f9fc5 | 4258 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
4259 | |
4260 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
4261 | if (!filter) { |
4262 | return true; // normal case | |
4263 | } | |
4264 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
4265 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
4266 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
4267 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
4268 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
4269 | return true; | |
31f18b77 | 4270 | } |
c07f9fc5 | 4271 | return false; |
31f18b77 FG |
4272 | } |
4273 | ||
4274 | bool should_dump_empty_bucket() const override { | |
4275 | return !filter; | |
4276 | } | |
7c673cae | 4277 | |
11fdf7f2 | 4278 | void init_table(TextTable *tbl) { |
7c673cae | 4279 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 4280 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
4281 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
4282 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
c07f9fc5 | 4283 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae | 4284 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 4285 | tbl->define_column("PRI-AFF", TextTable::LEFT, TextTable::RIGHT); |
11fdf7f2 TL |
4286 | } |
4287 | void dump(TextTable *tbl, string& bucket) { | |
4288 | init_table(tbl); | |
7c673cae | 4289 | |
11fdf7f2 TL |
4290 | if (!bucket.empty()) { |
4291 | set_root(bucket); | |
4292 | Parent::dump(tbl); | |
4293 | } else { | |
4294 | Parent::dump(tbl); | |
4295 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4296 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) { | |
4297 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), tbl); | |
4298 | } | |
31f18b77 | 4299 | } |
7c673cae FG |
4300 | } |
4301 | } | |
4302 | ||
4303 | protected: | |
4304 | void dump_item(const CrushTreeDumper::Item &qi, TextTable *tbl) override { | |
224ce89b WB |
4305 | const char *c = crush->get_item_class(qi.id); |
4306 | if (!c) | |
4307 | c = ""; | |
7c673cae | 4308 | *tbl << qi.id |
224ce89b | 4309 | << c |
7c673cae FG |
4310 | << weightf_t(qi.weight); |
4311 | ||
4312 | ostringstream name; | |
4313 | for (int k = 0; k < qi.depth; k++) | |
4314 | name << " "; | |
4315 | if (qi.is_bucket()) { | |
4316 | name << crush->get_type_name(crush->get_bucket_type(qi.id)) << " " | |
4317 | << crush->get_item_name(qi.id); | |
4318 | } else { | |
4319 | name << "osd." << qi.id; | |
4320 | } | |
4321 | *tbl << name.str(); | |
4322 | ||
4323 | if (!qi.is_bucket()) { | |
4324 | if (!osdmap->exists(qi.id)) { | |
4325 | *tbl << "DNE" | |
4326 | << 0; | |
4327 | } else { | |
c07f9fc5 FG |
4328 | string s; |
4329 | if (osdmap->is_up(qi.id)) { | |
4330 | s = "up"; | |
4331 | } else if (osdmap->is_destroyed(qi.id)) { | |
4332 | s = "destroyed"; | |
4333 | } else { | |
4334 | s = "down"; | |
4335 | } | |
4336 | *tbl << s | |
7c673cae FG |
4337 | << weightf_t(osdmap->get_weightf(qi.id)) |
4338 | << weightf_t(osdmap->get_primary_affinityf(qi.id)); | |
4339 | } | |
4340 | } | |
4341 | *tbl << TextTable::endrow; | |
4342 | } | |
4343 | ||
4344 | private: | |
4345 | const OSDMap *osdmap; | |
31f18b77 | 4346 | const unsigned filter; |
7c673cae FG |
4347 | }; |
4348 | ||
4349 | class OSDTreeFormattingDumper : public CrushTreeDumper::FormattingDumper { | |
4350 | public: | |
4351 | typedef CrushTreeDumper::FormattingDumper Parent; | |
4352 | ||
31f18b77 FG |
4353 | OSDTreeFormattingDumper(const CrushWrapper *crush, const OSDMap *osdmap_, |
4354 | unsigned f) | |
c07f9fc5 | 4355 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
4356 | |
4357 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
4358 | if (!filter) { |
4359 | return true; // normal case | |
4360 | } | |
4361 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
4362 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
4363 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
4364 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
4365 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
4366 | return true; | |
31f18b77 | 4367 | } |
c07f9fc5 | 4368 | return false; |
31f18b77 FG |
4369 | } |
4370 | ||
4371 | bool should_dump_empty_bucket() const override { | |
4372 | return !filter; | |
4373 | } | |
7c673cae | 4374 | |
11fdf7f2 TL |
4375 | void dump(Formatter *f, string& bucket) { |
4376 | if (!bucket.empty()) { | |
4377 | set_root(bucket); | |
4378 | f->open_array_section("nodes"); | |
4379 | Parent::dump(f); | |
4380 | f->close_section(); | |
4381 | } else { | |
4382 | f->open_array_section("nodes"); | |
4383 | Parent::dump(f); | |
4384 | f->close_section(); | |
4385 | f->open_array_section("stray"); | |
4386 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4387 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) | |
4388 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); | |
4389 | } | |
4390 | f->close_section(); | |
7c673cae | 4391 | } |
7c673cae FG |
4392 | } |
4393 | ||
4394 | protected: | |
4395 | void dump_item_fields(const CrushTreeDumper::Item &qi, Formatter *f) override { | |
4396 | Parent::dump_item_fields(qi, f); | |
4397 | if (!qi.is_bucket()) | |
4398 | { | |
c07f9fc5 FG |
4399 | string s; |
4400 | if (osdmap->is_up(qi.id)) { | |
4401 | s = "up"; | |
4402 | } else if (osdmap->is_destroyed(qi.id)) { | |
4403 | s = "destroyed"; | |
4404 | } else { | |
4405 | s = "down"; | |
4406 | } | |
7c673cae | 4407 | f->dump_unsigned("exists", (int)osdmap->exists(qi.id)); |
c07f9fc5 | 4408 | f->dump_string("status", s); |
7c673cae FG |
4409 | f->dump_float("reweight", osdmap->get_weightf(qi.id)); |
4410 | f->dump_float("primary_affinity", osdmap->get_primary_affinityf(qi.id)); | |
4411 | } | |
4412 | } | |
4413 | ||
4414 | private: | |
4415 | const OSDMap *osdmap; | |
31f18b77 | 4416 | const unsigned filter; |
7c673cae FG |
4417 | }; |
4418 | ||
11fdf7f2 | 4419 | void OSDMap::print_tree(Formatter *f, ostream *out, unsigned filter, string bucket) const |
7c673cae | 4420 | { |
31f18b77 | 4421 | if (f) { |
11fdf7f2 | 4422 | OSDTreeFormattingDumper(crush.get(), this, filter).dump(f, bucket); |
31f18b77 | 4423 | } else { |
11fdf7f2 | 4424 | ceph_assert(out); |
7c673cae | 4425 | TextTable tbl; |
11fdf7f2 | 4426 | OSDTreePlainDumper(crush.get(), this, filter).dump(&tbl, bucket); |
7c673cae FG |
4427 | *out << tbl; |
4428 | } | |
4429 | } | |
4430 | ||
224ce89b | 4431 | void OSDMap::print_summary(Formatter *f, ostream& out, |
11fdf7f2 | 4432 | const string& prefix, bool extra) const |
7c673cae FG |
4433 | { |
4434 | if (f) { | |
7c673cae FG |
4435 | f->dump_int("epoch", get_epoch()); |
4436 | f->dump_int("num_osds", get_num_osds()); | |
4437 | f->dump_int("num_up_osds", get_num_up_osds()); | |
9f95a23c | 4438 | f->dump_int("osd_up_since", last_up_change.to_msec() / 1000); |
7c673cae | 4439 | f->dump_int("num_in_osds", get_num_in_osds()); |
9f95a23c | 4440 | f->dump_int("osd_in_since", last_in_change.to_msec() / 1000); |
7c673cae | 4441 | f->dump_unsigned("num_remapped_pgs", get_num_pg_temp()); |
7c673cae | 4442 | } else { |
11fdf7f2 | 4443 | utime_t now = ceph_clock_now(); |
31f18b77 | 4444 | out << get_num_osds() << " osds: " |
11fdf7f2 TL |
4445 | << get_num_up_osds() << " up"; |
4446 | if (last_up_change != utime_t()) { | |
4447 | out << " (since " << utimespan_str(now - last_up_change) << ")"; | |
4448 | } | |
4449 | out << ", " << get_num_in_osds() << " in"; | |
4450 | if (last_in_change != utime_t()) { | |
4451 | out << " (since " << utimespan_str(now - last_in_change) << ")"; | |
4452 | } | |
4453 | if (extra) | |
4454 | out << "; epoch: e" << get_epoch(); | |
7c673cae FG |
4455 | if (get_num_pg_temp()) |
4456 | out << "; " << get_num_pg_temp() << " remapped pgs"; | |
4457 | out << "\n"; | |
4458 | uint64_t important_flags = flags & ~CEPH_OSDMAP_SEMIHIDDEN_FLAGS; | |
4459 | if (important_flags) | |
224ce89b | 4460 | out << prefix << "flags " << get_flag_string(important_flags) << "\n"; |
7c673cae FG |
4461 | } |
4462 | } | |
4463 | ||
4464 | void OSDMap::print_oneline_summary(ostream& out) const | |
4465 | { | |
4466 | out << "e" << get_epoch() << ": " | |
31f18b77 | 4467 | << get_num_osds() << " total, " |
7c673cae FG |
4468 | << get_num_up_osds() << " up, " |
4469 | << get_num_in_osds() << " in"; | |
7c673cae FG |
4470 | } |
4471 | ||
3efd9988 | 4472 | bool OSDMap::crush_rule_in_use(int rule_id) const |
7c673cae FG |
4473 | { |
4474 | for (const auto &pool : pools) { | |
3efd9988 | 4475 | if (pool.second.crush_rule == rule_id) |
7c673cae FG |
4476 | return true; |
4477 | } | |
4478 | return false; | |
4479 | } | |
4480 | ||
3efd9988 FG |
4481 | int OSDMap::validate_crush_rules(CrushWrapper *newcrush, |
4482 | ostream *ss) const | |
4483 | { | |
4484 | for (auto& i : pools) { | |
4485 | auto& pool = i.second; | |
4486 | int ruleno = pool.get_crush_rule(); | |
4487 | if (!newcrush->rule_exists(ruleno)) { | |
4488 | *ss << "pool " << i.first << " references crush_rule " << ruleno | |
4489 | << " but it is not present"; | |
4490 | return -EINVAL; | |
4491 | } | |
20effc67 | 4492 | if (newcrush->get_rule_type(ruleno) != (int)pool.get_type()) { |
3efd9988 FG |
4493 | *ss << "pool " << i.first << " type does not match rule " << ruleno; |
4494 | return -EINVAL; | |
4495 | } | |
3efd9988 FG |
4496 | } |
4497 | return 0; | |
4498 | } | |
4499 | ||
224ce89b WB |
4500 | int OSDMap::build_simple_optioned(CephContext *cct, epoch_t e, uuid_d &fsid, |
4501 | int nosd, int pg_bits, int pgp_bits, | |
4502 | bool default_pool) | |
7c673cae | 4503 | { |
224ce89b WB |
4504 | ldout(cct, 10) << "build_simple on " << nosd |
4505 | << " osds" << dendl; | |
7c673cae FG |
4506 | epoch = e; |
4507 | set_fsid(fsid); | |
4508 | created = modified = ceph_clock_now(); | |
4509 | ||
4510 | if (nosd >= 0) { | |
4511 | set_max_osd(nosd); | |
4512 | } else { | |
4513 | // count osds | |
4514 | int maxosd = 0; | |
11fdf7f2 | 4515 | const auto& conf = cct->_conf; |
7c673cae | 4516 | vector<string> sections; |
11fdf7f2 | 4517 | conf.get_all_sections(sections); |
7c673cae FG |
4518 | |
4519 | for (auto §ion : sections) { | |
4520 | if (section.find("osd.") != 0) | |
4521 | continue; | |
4522 | ||
4523 | const char *begin = section.c_str() + 4; | |
4524 | char *end = (char*)begin; | |
4525 | int o = strtol(begin, &end, 10); | |
4526 | if (*end != '\0') | |
4527 | continue; | |
4528 | ||
4529 | if (o > cct->_conf->mon_max_osd) { | |
4530 | lderr(cct) << "[osd." << o << "] in config has id > mon_max_osd " << cct->_conf->mon_max_osd << dendl; | |
4531 | return -ERANGE; | |
4532 | } | |
4533 | ||
4534 | if (o > maxosd) | |
4535 | maxosd = o; | |
4536 | } | |
4537 | ||
4538 | set_max_osd(maxosd + 1); | |
4539 | } | |
4540 | ||
7c673cae FG |
4541 | |
4542 | stringstream ss; | |
4543 | int r; | |
4544 | if (nosd >= 0) | |
4545 | r = build_simple_crush_map(cct, *crush, nosd, &ss); | |
4546 | else | |
4547 | r = build_simple_crush_map_from_conf(cct, *crush, &ss); | |
11fdf7f2 | 4548 | ceph_assert(r == 0); |
7c673cae FG |
4549 | |
4550 | int poolbase = get_max_osd() ? get_max_osd() : 1; | |
4551 | ||
20effc67 | 4552 | const int default_replicated_rule = crush->get_osd_pool_default_crush_replicated_rule(cct); |
11fdf7f2 | 4553 | ceph_assert(default_replicated_rule >= 0); |
7c673cae | 4554 | |
224ce89b WB |
4555 | if (default_pool) { |
4556 | // pgp_num <= pg_num | |
4557 | if (pgp_bits > pg_bits) | |
4558 | pgp_bits = pg_bits; | |
4559 | ||
4560 | vector<string> pool_names; | |
4561 | pool_names.push_back("rbd"); | |
4562 | for (auto &plname : pool_names) { | |
4563 | int64_t pool = ++pool_max; | |
4564 | pools[pool].type = pg_pool_t::TYPE_REPLICATED; | |
4565 | pools[pool].flags = cct->_conf->osd_pool_default_flags; | |
4566 | if (cct->_conf->osd_pool_default_flag_hashpspool) | |
4567 | pools[pool].set_flag(pg_pool_t::FLAG_HASHPSPOOL); | |
4568 | if (cct->_conf->osd_pool_default_flag_nodelete) | |
4569 | pools[pool].set_flag(pg_pool_t::FLAG_NODELETE); | |
4570 | if (cct->_conf->osd_pool_default_flag_nopgchange) | |
4571 | pools[pool].set_flag(pg_pool_t::FLAG_NOPGCHANGE); | |
4572 | if (cct->_conf->osd_pool_default_flag_nosizechange) | |
4573 | pools[pool].set_flag(pg_pool_t::FLAG_NOSIZECHANGE); | |
20effc67 TL |
4574 | if (cct->_conf->osd_pool_default_flag_bulk) |
4575 | pools[pool].set_flag(pg_pool_t::FLAG_BULK); | |
11fdf7f2 TL |
4576 | pools[pool].size = cct->_conf.get_val<uint64_t>("osd_pool_default_size"); |
4577 | pools[pool].min_size = cct->_conf.get_osd_pool_default_min_size( | |
4578 | pools[pool].size); | |
224ce89b WB |
4579 | pools[pool].crush_rule = default_replicated_rule; |
4580 | pools[pool].object_hash = CEPH_STR_HASH_RJENKINS; | |
4581 | pools[pool].set_pg_num(poolbase << pg_bits); | |
4582 | pools[pool].set_pgp_num(poolbase << pgp_bits); | |
11fdf7f2 TL |
4583 | pools[pool].set_pg_num_target(poolbase << pg_bits); |
4584 | pools[pool].set_pgp_num_target(poolbase << pgp_bits); | |
224ce89b | 4585 | pools[pool].last_change = epoch; |
c07f9fc5 FG |
4586 | pools[pool].application_metadata.insert( |
4587 | {pg_pool_t::APPLICATION_NAME_RBD, {}}); | |
9f95a23c TL |
4588 | if (auto m = pg_pool_t::get_pg_autoscale_mode_by_name( |
4589 | cct->_conf.get_val<string>("osd_pool_default_pg_autoscale_mode")); | |
4590 | m != pg_pool_t::pg_autoscale_mode_t::UNKNOWN) { | |
4591 | pools[pool].pg_autoscale_mode = m; | |
4592 | } else { | |
4593 | pools[pool].pg_autoscale_mode = pg_pool_t::pg_autoscale_mode_t::OFF; | |
4594 | } | |
224ce89b WB |
4595 | pool_name[pool] = plname; |
4596 | name_pool[plname] = pool; | |
4597 | } | |
7c673cae FG |
4598 | } |
4599 | ||
7c673cae FG |
4600 | map<string,string> profile_map; |
4601 | r = get_erasure_code_profile_default(cct, profile_map, &ss); | |
4602 | if (r < 0) { | |
4603 | lderr(cct) << ss.str() << dendl; | |
4604 | return r; | |
4605 | } | |
4606 | set_erasure_code_profile("default", profile_map); | |
4607 | return 0; | |
4608 | } | |
4609 | ||
4610 | int OSDMap::get_erasure_code_profile_default(CephContext *cct, | |
4611 | map<string,string> &profile_map, | |
4612 | ostream *ss) | |
4613 | { | |
11fdf7f2 | 4614 | int r = get_json_str_map(cct->_conf.get_val<string>("osd_pool_default_erasure_code_profile"), |
7c673cae FG |
4615 | *ss, |
4616 | &profile_map); | |
4617 | return r; | |
4618 | } | |
4619 | ||
4620 | int OSDMap::_build_crush_types(CrushWrapper& crush) | |
4621 | { | |
4622 | crush.set_type_name(0, "osd"); | |
4623 | crush.set_type_name(1, "host"); | |
4624 | crush.set_type_name(2, "chassis"); | |
4625 | crush.set_type_name(3, "rack"); | |
4626 | crush.set_type_name(4, "row"); | |
4627 | crush.set_type_name(5, "pdu"); | |
4628 | crush.set_type_name(6, "pod"); | |
4629 | crush.set_type_name(7, "room"); | |
4630 | crush.set_type_name(8, "datacenter"); | |
11fdf7f2 TL |
4631 | crush.set_type_name(9, "zone"); |
4632 | crush.set_type_name(10, "region"); | |
4633 | crush.set_type_name(11, "root"); | |
4634 | return 11; | |
7c673cae FG |
4635 | } |
4636 | ||
4637 | int OSDMap::build_simple_crush_map(CephContext *cct, CrushWrapper& crush, | |
4638 | int nosd, ostream *ss) | |
4639 | { | |
4640 | crush.create(); | |
4641 | ||
4642 | // root | |
4643 | int root_type = _build_crush_types(crush); | |
4644 | int rootid; | |
4645 | int r = crush.add_bucket(0, 0, CRUSH_HASH_DEFAULT, | |
4646 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4647 | ceph_assert(r == 0); |
7c673cae FG |
4648 | crush.set_item_name(rootid, "default"); |
4649 | ||
f67539c2 TL |
4650 | map<string,string> loc{ |
4651 | {"host", "localhost"}, | |
4652 | {"rack", "localrack"}, | |
4653 | {"root", "default"} | |
4654 | }; | |
7c673cae | 4655 | for (int o=0; o<nosd; o++) { |
7c673cae FG |
4656 | ldout(cct, 10) << " adding osd." << o << " at " << loc << dendl; |
4657 | char name[32]; | |
4658 | snprintf(name, sizeof(name), "osd.%d", o); | |
4659 | crush.insert_item(cct, o, 1.0, name, loc); | |
4660 | } | |
4661 | ||
31f18b77 | 4662 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4663 | |
4664 | crush.finalize(); | |
4665 | ||
4666 | return 0; | |
4667 | } | |
4668 | ||
4669 | int OSDMap::build_simple_crush_map_from_conf(CephContext *cct, | |
4670 | CrushWrapper& crush, | |
4671 | ostream *ss) | |
4672 | { | |
11fdf7f2 | 4673 | const auto& conf = cct->_conf; |
7c673cae FG |
4674 | |
4675 | crush.create(); | |
4676 | ||
4677 | // root | |
4678 | int root_type = _build_crush_types(crush); | |
4679 | int rootid; | |
4680 | int r = crush.add_bucket(0, 0, | |
4681 | CRUSH_HASH_DEFAULT, | |
4682 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4683 | ceph_assert(r == 0); |
7c673cae FG |
4684 | crush.set_item_name(rootid, "default"); |
4685 | ||
4686 | // add osds | |
4687 | vector<string> sections; | |
11fdf7f2 | 4688 | conf.get_all_sections(sections); |
7c673cae FG |
4689 | |
4690 | for (auto §ion : sections) { | |
4691 | if (section.find("osd.") != 0) | |
4692 | continue; | |
4693 | ||
4694 | const char *begin = section.c_str() + 4; | |
4695 | char *end = (char*)begin; | |
4696 | int o = strtol(begin, &end, 10); | |
4697 | if (*end != '\0') | |
4698 | continue; | |
4699 | ||
4700 | string host, rack, row, room, dc, pool; | |
4701 | vector<string> sectiontmp; | |
4702 | sectiontmp.push_back("osd"); | |
4703 | sectiontmp.push_back(section); | |
11fdf7f2 TL |
4704 | conf.get_val_from_conf_file(sectiontmp, "host", host, false); |
4705 | conf.get_val_from_conf_file(sectiontmp, "rack", rack, false); | |
4706 | conf.get_val_from_conf_file(sectiontmp, "row", row, false); | |
4707 | conf.get_val_from_conf_file(sectiontmp, "room", room, false); | |
4708 | conf.get_val_from_conf_file(sectiontmp, "datacenter", dc, false); | |
4709 | conf.get_val_from_conf_file(sectiontmp, "root", pool, false); | |
7c673cae FG |
4710 | |
4711 | if (host.length() == 0) | |
4712 | host = "unknownhost"; | |
4713 | if (rack.length() == 0) | |
4714 | rack = "unknownrack"; | |
4715 | ||
4716 | map<string,string> loc; | |
4717 | loc["host"] = host; | |
4718 | loc["rack"] = rack; | |
4719 | if (row.size()) | |
4720 | loc["row"] = row; | |
4721 | if (room.size()) | |
4722 | loc["room"] = room; | |
4723 | if (dc.size()) | |
4724 | loc["datacenter"] = dc; | |
4725 | loc["root"] = "default"; | |
4726 | ||
4727 | ldout(cct, 5) << " adding osd." << o << " at " << loc << dendl; | |
4728 | crush.insert_item(cct, o, 1.0, section, loc); | |
4729 | } | |
4730 | ||
31f18b77 | 4731 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4732 | |
4733 | crush.finalize(); | |
4734 | ||
4735 | return 0; | |
4736 | } | |
4737 | ||
4738 | ||
31f18b77 FG |
4739 | int OSDMap::build_simple_crush_rules( |
4740 | CephContext *cct, | |
4741 | CrushWrapper& crush, | |
4742 | const string& root, | |
4743 | ostream *ss) | |
7c673cae | 4744 | { |
20effc67 | 4745 | int crush_rule = crush.get_osd_pool_default_crush_replicated_rule(cct); |
7c673cae FG |
4746 | string failure_domain = |
4747 | crush.get_type_name(cct->_conf->osd_crush_chooseleaf_type); | |
4748 | ||
7c673cae | 4749 | int r; |
31f18b77 | 4750 | r = crush.add_simple_rule_at( |
224ce89b | 4751 | "replicated_rule", root, failure_domain, "", |
31f18b77 FG |
4752 | "firstn", pg_pool_t::TYPE_REPLICATED, |
4753 | crush_rule, ss); | |
7c673cae FG |
4754 | if (r < 0) |
4755 | return r; | |
4756 | // do not add an erasure rule by default or else we will implicitly | |
4757 | // require the crush_v2 feature of clients | |
4758 | return 0; | |
4759 | } | |
4760 | ||
4761 | int OSDMap::summarize_mapping_stats( | |
4762 | OSDMap *newmap, | |
4763 | const set<int64_t> *pools, | |
4764 | std::string *out, | |
4765 | Formatter *f) const | |
4766 | { | |
4767 | set<int64_t> ls; | |
4768 | if (pools) { | |
4769 | ls = *pools; | |
4770 | } else { | |
4771 | for (auto &p : get_pools()) | |
4772 | ls.insert(p.first); | |
4773 | } | |
4774 | ||
4775 | unsigned total_pg = 0; | |
4776 | unsigned moved_pg = 0; | |
4777 | vector<unsigned> base_by_osd(get_max_osd(), 0); | |
4778 | vector<unsigned> new_by_osd(get_max_osd(), 0); | |
4779 | for (int64_t pool_id : ls) { | |
4780 | const pg_pool_t *pi = get_pg_pool(pool_id); | |
31f18b77 FG |
4781 | vector<int> up, up2; |
4782 | int up_primary; | |
7c673cae | 4783 | for (unsigned ps = 0; ps < pi->get_pg_num(); ++ps) { |
11fdf7f2 | 4784 | pg_t pgid(ps, pool_id); |
7c673cae | 4785 | total_pg += pi->get_size(); |
31f18b77 | 4786 | pg_to_up_acting_osds(pgid, &up, &up_primary, nullptr, nullptr); |
7c673cae FG |
4787 | for (int osd : up) { |
4788 | if (osd >= 0 && osd < get_max_osd()) | |
4789 | ++base_by_osd[osd]; | |
4790 | } | |
4791 | if (newmap) { | |
31f18b77 | 4792 | newmap->pg_to_up_acting_osds(pgid, &up2, &up_primary, nullptr, nullptr); |
7c673cae FG |
4793 | for (int osd : up2) { |
4794 | if (osd >= 0 && osd < get_max_osd()) | |
4795 | ++new_by_osd[osd]; | |
4796 | } | |
1e59de90 | 4797 | if (pi->is_erasure()) { |
7c673cae FG |
4798 | for (unsigned i=0; i<up.size(); ++i) { |
4799 | if (up[i] != up2[i]) { | |
4800 | ++moved_pg; | |
4801 | } | |
4802 | } | |
1e59de90 | 4803 | } else if (pi->is_replicated()) { |
7c673cae FG |
4804 | for (int osd : up) { |
4805 | if (std::find(up2.begin(), up2.end(), osd) == up2.end()) { | |
4806 | ++moved_pg; | |
4807 | } | |
4808 | } | |
4809 | } else { | |
11fdf7f2 | 4810 | ceph_abort_msg("unhandled pool type"); |
7c673cae FG |
4811 | } |
4812 | } | |
4813 | } | |
4814 | } | |
4815 | ||
4816 | unsigned num_up_in = 0; | |
4817 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4818 | if (is_up(osd) && is_in(osd)) | |
4819 | ++num_up_in; | |
4820 | } | |
4821 | if (!num_up_in) { | |
4822 | return -EINVAL; | |
4823 | } | |
4824 | ||
4825 | float avg_pg = (float)total_pg / (float)num_up_in; | |
4826 | float base_stddev = 0, new_stddev = 0; | |
4827 | int min = -1, max = -1; | |
4828 | unsigned min_base_pg = 0, max_base_pg = 0; | |
4829 | unsigned min_new_pg = 0, max_new_pg = 0; | |
4830 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4831 | if (is_up(osd) && is_in(osd)) { | |
4832 | float base_diff = (float)base_by_osd[osd] - avg_pg; | |
4833 | base_stddev += base_diff * base_diff; | |
4834 | float new_diff = (float)new_by_osd[osd] - avg_pg; | |
4835 | new_stddev += new_diff * new_diff; | |
4836 | if (min < 0 || base_by_osd[osd] < min_base_pg) { | |
4837 | min = osd; | |
4838 | min_base_pg = base_by_osd[osd]; | |
4839 | min_new_pg = new_by_osd[osd]; | |
4840 | } | |
4841 | if (max < 0 || base_by_osd[osd] > max_base_pg) { | |
4842 | max = osd; | |
4843 | max_base_pg = base_by_osd[osd]; | |
4844 | max_new_pg = new_by_osd[osd]; | |
4845 | } | |
4846 | } | |
4847 | } | |
4848 | base_stddev = sqrt(base_stddev / num_up_in); | |
4849 | new_stddev = sqrt(new_stddev / num_up_in); | |
4850 | ||
4851 | float edev = sqrt(avg_pg * (1.0 - (1.0 / (double)num_up_in))); | |
4852 | ||
4853 | ostringstream ss; | |
4854 | if (f) | |
4855 | f->open_object_section("utilization"); | |
4856 | if (newmap) { | |
4857 | if (f) { | |
4858 | f->dump_unsigned("moved_pgs", moved_pg); | |
4859 | f->dump_unsigned("total_pgs", total_pg); | |
4860 | } else { | |
4861 | float percent = 0; | |
4862 | if (total_pg) | |
4863 | percent = (float)moved_pg * 100.0 / (float)total_pg; | |
4864 | ss << "moved " << moved_pg << " / " << total_pg | |
4865 | << " (" << percent << "%)\n"; | |
4866 | } | |
4867 | } | |
4868 | if (f) { | |
4869 | f->dump_float("avg_pgs", avg_pg); | |
4870 | f->dump_float("std_dev", base_stddev); | |
4871 | f->dump_float("expected_baseline_std_dev", edev); | |
4872 | if (newmap) | |
4873 | f->dump_float("new_std_dev", new_stddev); | |
4874 | } else { | |
4875 | ss << "avg " << avg_pg << "\n"; | |
4876 | ss << "stddev " << base_stddev; | |
4877 | if (newmap) | |
4878 | ss << " -> " << new_stddev; | |
4879 | ss << " (expected baseline " << edev << ")\n"; | |
4880 | } | |
4881 | if (min >= 0) { | |
4882 | if (f) { | |
4883 | f->dump_unsigned("min_osd", min); | |
4884 | f->dump_unsigned("min_osd_pgs", min_base_pg); | |
4885 | if (newmap) | |
4886 | f->dump_unsigned("new_min_osd_pgs", min_new_pg); | |
4887 | } else { | |
4888 | ss << "min osd." << min << " with " << min_base_pg; | |
4889 | if (newmap) | |
4890 | ss << " -> " << min_new_pg; | |
4891 | ss << " pgs (" << (float)min_base_pg / avg_pg; | |
4892 | if (newmap) | |
4893 | ss << " -> " << (float)min_new_pg / avg_pg; | |
4894 | ss << " * mean)\n"; | |
4895 | } | |
4896 | } | |
4897 | if (max >= 0) { | |
4898 | if (f) { | |
4899 | f->dump_unsigned("max_osd", max); | |
4900 | f->dump_unsigned("max_osd_pgs", max_base_pg); | |
4901 | if (newmap) | |
4902 | f->dump_unsigned("new_max_osd_pgs", max_new_pg); | |
4903 | } else { | |
4904 | ss << "max osd." << max << " with " << max_base_pg; | |
4905 | if (newmap) | |
4906 | ss << " -> " << max_new_pg; | |
4907 | ss << " pgs (" << (float)max_base_pg / avg_pg; | |
4908 | if (newmap) | |
4909 | ss << " -> " << (float)max_new_pg / avg_pg; | |
4910 | ss << " * mean)\n"; | |
4911 | } | |
4912 | } | |
4913 | if (f) | |
4914 | f->close_section(); | |
4915 | if (out) | |
4916 | *out = ss.str(); | |
4917 | return 0; | |
4918 | } | |
4919 | ||
7c673cae FG |
4920 | bool OSDMap::try_pg_upmap( |
4921 | CephContext *cct, | |
4922 | pg_t pg, ///< pg to potentially remap | |
4923 | const set<int>& overfull, ///< osds we'd want to evacuate | |
4924 | const vector<int>& underfull, ///< osds to move to, in order of preference | |
92f5a8d4 | 4925 | const vector<int>& more_underfull, ///< more osds only slightly underfull |
7c673cae FG |
4926 | vector<int> *orig, |
4927 | vector<int> *out) ///< resulting alternative mapping | |
4928 | { | |
4929 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
4930 | if (!pool) | |
4931 | return false; | |
20effc67 | 4932 | int rule = pool->get_crush_rule(); |
7c673cae FG |
4933 | if (rule < 0) |
4934 | return false; | |
4935 | ||
7c673cae FG |
4936 | // make sure there is something there to remap |
4937 | bool any = false; | |
4938 | for (auto osd : *orig) { | |
4939 | if (overfull.count(osd)) { | |
4940 | any = true; | |
4941 | break; | |
4942 | } | |
4943 | } | |
4944 | if (!any) { | |
4945 | return false; | |
4946 | } | |
4947 | ||
4948 | int r = crush->try_remap_rule( | |
4949 | cct, | |
4950 | rule, | |
4951 | pool->get_size(), | |
4952 | overfull, underfull, | |
92f5a8d4 | 4953 | more_underfull, |
7c673cae FG |
4954 | *orig, |
4955 | out); | |
4956 | if (r < 0) | |
4957 | return false; | |
4958 | if (*out == *orig) | |
4959 | return false; | |
4960 | return true; | |
4961 | } | |
4962 | ||
1e59de90 TL |
4963 | |
4964 | int OSDMap::balance_primaries( | |
4965 | CephContext *cct, | |
4966 | int64_t pid, | |
4967 | OSDMap::Incremental *pending_inc, | |
4968 | OSDMap& tmp_osd_map) const | |
4969 | { | |
4970 | // This function only handles replicated pools. | |
4971 | const pg_pool_t* pool = get_pg_pool(pid); | |
4972 | if (! pool->is_replicated()) { | |
4973 | ldout(cct, 10) << __func__ << " skipping erasure pool " | |
4974 | << get_pool_name(pid) << dendl; | |
4975 | return -EINVAL; | |
4976 | } | |
4977 | ||
4978 | // Info to be used in verify_upmap | |
4979 | int pool_size = pool->get_size(); | |
4980 | int crush_rule = pool->get_crush_rule(); | |
4981 | ||
4982 | // Get pgs by osd (map of osd -> pgs) | |
4983 | // Get primaries by osd (map of osd -> primary) | |
4984 | map<uint64_t,set<pg_t>> pgs_by_osd; | |
4985 | map<uint64_t,set<pg_t>> prim_pgs_by_osd; | |
4986 | map<uint64_t,set<pg_t>> acting_prims_by_osd; | |
4987 | pgs_by_osd = tmp_osd_map.get_pgs_by_osd(cct, pid, &prim_pgs_by_osd, &acting_prims_by_osd); | |
4988 | ||
aee94f69 TL |
4989 | // Construct information about the pgs and osds we will consider in new primary mappings, |
4990 | // as well as a map of all pgs and their original primary osds. | |
1e59de90 TL |
4991 | map<pg_t,bool> prim_pgs_to_check; |
4992 | vector<uint64_t> osds_to_check; | |
aee94f69 | 4993 | map<pg_t, uint64_t> orig_prims; |
1e59de90 TL |
4994 | for (const auto & [osd, pgs] : prim_pgs_by_osd) { |
4995 | osds_to_check.push_back(osd); | |
4996 | for (const auto & pg : pgs) { | |
4997 | prim_pgs_to_check.insert({pg, false}); | |
aee94f69 | 4998 | orig_prims.insert({pg, osd}); |
1e59de90 TL |
4999 | } |
5000 | } | |
5001 | ||
5002 | // calculate desired primary distribution for each osd | |
5003 | map<uint64_t,float> desired_prim_dist; | |
5004 | int rc = 0; | |
5005 | rc = calc_desired_primary_distribution(cct, pid, osds_to_check, desired_prim_dist); | |
5006 | if (rc < 0) { | |
5007 | ldout(cct, 10) << __func__ << " Error in calculating desired primary distribution" << dendl; | |
5008 | return -EINVAL; | |
5009 | } | |
5010 | map<uint64_t,float> prim_dist_scores; | |
5011 | float actual; | |
5012 | float desired; | |
5013 | for (auto osd : osds_to_check) { | |
5014 | actual = prim_pgs_by_osd[osd].size(); | |
5015 | desired = desired_prim_dist[osd]; | |
5016 | prim_dist_scores[osd] = actual - desired; | |
5017 | ldout(cct, 10) << __func__ << " desired distribution for osd." << osd << " " << desired << dendl; | |
5018 | } | |
5019 | ||
5020 | // get read balance score before balancing | |
5021 | float read_balance_score_before = 0.0; | |
5022 | read_balance_info_t rb_info; | |
5023 | rc = tmp_osd_map.calc_read_balance_score(cct, pid, &rb_info); | |
5024 | if (rc >= 0) { | |
5025 | read_balance_score_before = rb_info.adjusted_score; | |
5026 | } | |
5027 | if (rb_info.err_msg.length() > 0) { | |
5028 | ldout(cct, 10) << __func__ << (rc < 0 ? " ERROR: " : " Warning: ") << rb_info.err_msg << dendl; | |
5029 | return -EINVAL; | |
5030 | } | |
5031 | ||
5032 | // get ready to swap pgs | |
5033 | while (true) { | |
5034 | int curr_num_changes = 0; | |
5035 | vector<int> up_osds; | |
5036 | vector<int> acting_osds; | |
5037 | int up_primary, acting_primary; | |
5038 | for (const auto & [pg, mapped] : prim_pgs_to_check) { | |
5039 | // fill in the up, up primary, acting, and acting primary for the current PG | |
5040 | tmp_osd_map.pg_to_up_acting_osds(pg, &up_osds, &up_primary, | |
5041 | &acting_osds, &acting_primary); | |
5042 | ||
5043 | // find the OSD that would make the best swap based on its score | |
5044 | // We start by first testing the OSD that is currently primary for the PG we are checking. | |
5045 | uint64_t curr_best_osd = up_primary; | |
5046 | float prim_score = prim_dist_scores[up_primary]; | |
5047 | for (auto potential_osd : up_osds) { | |
5048 | float potential_score = prim_dist_scores[potential_osd]; | |
5049 | if ((prim_score > 0) && // taking 1 pg from the prim would not make its score worse | |
5050 | (potential_score < 0) && // adding 1 pg to the potential would not make its score worse | |
5051 | ((prim_score - potential_score) > 1) && // swapping a pg would not just keep the scores the same | |
5052 | (desired_prim_dist[potential_osd] > 0)) // the potential is not off limits (the primary affinity is above 0) | |
5053 | { | |
5054 | curr_best_osd = potential_osd; | |
5055 | } | |
5056 | } | |
5057 | ||
5058 | // Make the swap only if: | |
5059 | // 1. The swap is legal | |
5060 | // 2. The balancer has chosen a new primary | |
5061 | auto legal_swap = crush->verify_upmap(cct, | |
5062 | crush_rule, | |
5063 | pool_size, | |
5064 | {(int)curr_best_osd}); | |
5065 | if (legal_swap >= 0 && | |
5066 | ((int)curr_best_osd != up_primary)) { | |
5067 | // Update prim_dist_scores | |
5068 | prim_dist_scores[curr_best_osd] += 1; | |
5069 | prim_dist_scores[up_primary] -= 1; | |
5070 | ||
5071 | // Update the mappings | |
1e59de90 | 5072 | tmp_osd_map.pg_upmap_primaries[pg] = curr_best_osd; |
aee94f69 TL |
5073 | if (curr_best_osd == orig_prims[pg]) { |
5074 | pending_inc->new_pg_upmap_primary.erase(pg); | |
5075 | prim_pgs_to_check[pg] = false; | |
5076 | } else { | |
5077 | pending_inc->new_pg_upmap_primary[pg] = curr_best_osd; | |
5078 | prim_pgs_to_check[pg] = true; // mark that this pg changed mappings | |
5079 | } | |
1e59de90 TL |
5080 | |
5081 | curr_num_changes++; | |
5082 | } | |
5083 | ldout(cct, 20) << __func__ << " curr_num_changes: " << curr_num_changes << dendl; | |
5084 | } | |
5085 | // If there are no changes after one pass through the pgs, then no further optimizations can be made. | |
5086 | if (curr_num_changes == 0) { | |
5087 | ldout(cct, 20) << __func__ << " curr_num_changes is 0; no further optimizations can be made." << dendl; | |
5088 | break; | |
5089 | } | |
5090 | } | |
5091 | ||
5092 | // get read balance score after balancing | |
5093 | float read_balance_score_after = 0.0; | |
5094 | rc = tmp_osd_map.calc_read_balance_score(cct, pid, &rb_info); | |
5095 | if (rc >= 0) { | |
5096 | read_balance_score_after = rb_info.adjusted_score; | |
5097 | } | |
5098 | if (rb_info.err_msg.length() > 0) { | |
5099 | ldout(cct, 10) << __func__ << (rc < 0 ? " ERROR: " : " Warning: ") << rb_info.err_msg << dendl; | |
5100 | return -EINVAL; | |
5101 | } | |
5102 | ||
5103 | // Tally total number of changes | |
5104 | int num_changes = 0; | |
5105 | if (read_balance_score_after < read_balance_score_before) { | |
5106 | for (auto [pg, mapped] : prim_pgs_to_check) { | |
5107 | if (mapped) { | |
5108 | num_changes++; | |
5109 | } | |
5110 | } | |
5111 | } | |
5112 | ||
5113 | ldout(cct, 10) << __func__ << " num_changes " << num_changes << dendl; | |
5114 | return num_changes; | |
5115 | } | |
5116 | ||
5117 | int OSDMap::calc_desired_primary_distribution( | |
5118 | CephContext *cct, | |
5119 | int64_t pid, | |
5120 | const vector<uint64_t> &osds, | |
5121 | std::map<uint64_t, float>& desired_primary_distribution) const | |
5122 | { | |
5123 | // will return a perfect distribution of floats | |
5124 | // without calculating the floor of each value | |
5125 | // | |
5126 | // This function only handles replicated pools. | |
5127 | const pg_pool_t* pool = get_pg_pool(pid); | |
5128 | if (pool->is_replicated()) { | |
5129 | ldout(cct, 20) << __func__ << " calculating distribution for replicated pool " | |
5130 | << get_pool_name(pid) << dendl; | |
5131 | uint64_t replica_count = pool->get_size(); | |
5132 | ||
5133 | map<uint64_t,set<pg_t>> pgs_by_osd; | |
5134 | pgs_by_osd = get_pgs_by_osd(cct, pid); | |
5135 | ||
5136 | // First calculate the distribution using primary affinity and tally up the sum | |
5137 | auto distribution_sum = 0.0; | |
5138 | for (const auto & osd : osds) { | |
5139 | float osd_primary_count = ((float)pgs_by_osd[osd].size() / (float)replica_count) * get_primary_affinityf(osd); | |
5140 | desired_primary_distribution.insert({osd, osd_primary_count}); | |
5141 | distribution_sum += osd_primary_count; | |
5142 | } | |
5143 | if (distribution_sum <= 0) { | |
5144 | ldout(cct, 10) << __func__ << " Unable to calculate primary distribution, likely because primary affinity is" | |
5145 | << " set to 0 on all OSDs." << dendl; | |
5146 | return -EINVAL; | |
5147 | } | |
5148 | ||
5149 | // Then, stretch the value (necessary when primary affinity is smaller than 1) | |
5150 | float factor = (float)pool->get_pg_num() / (float)distribution_sum; | |
5151 | float distribution_sum_desired = 0.0; | |
5152 | ||
5153 | ceph_assert(factor >= 1.0); | |
5154 | for (const auto & [osd, osd_primary_count] : desired_primary_distribution) { | |
5155 | desired_primary_distribution[osd] *= factor; | |
5156 | distribution_sum_desired += desired_primary_distribution[osd]; | |
5157 | } | |
5158 | ceph_assert(fabs(distribution_sum_desired - pool->get_pg_num()) < 0.01); | |
5159 | } else { | |
5160 | ldout(cct, 10) << __func__ <<" skipping erasure pool " | |
5161 | << get_pool_name(pid) << dendl; | |
5162 | return -EINVAL; | |
5163 | } | |
5164 | ||
5165 | return 0; | |
5166 | } | |
5167 | ||
7c673cae FG |
5168 | int OSDMap::calc_pg_upmaps( |
5169 | CephContext *cct, | |
92f5a8d4 | 5170 | uint32_t max_deviation, |
7c673cae | 5171 | int max, |
a8e16298 | 5172 | const set<int64_t>& only_pools, |
20effc67 TL |
5173 | OSDMap::Incremental *pending_inc, |
5174 | std::random_device::result_type *p_seed) | |
7c673cae | 5175 | { |
a8e16298 | 5176 | ldout(cct, 10) << __func__ << " pools " << only_pools << dendl; |
20effc67 | 5177 | OSDMap tmp_osd_map; |
92f5a8d4 TL |
5178 | // Can't be less than 1 pg |
5179 | if (max_deviation < 1) | |
5180 | max_deviation = 1; | |
20effc67 | 5181 | tmp_osd_map.deepish_copy_from(*this); |
7c673cae | 5182 | int num_changed = 0; |
a8e16298 TL |
5183 | map<int,set<pg_t>> pgs_by_osd; |
5184 | int total_pgs = 0; | |
5185 | float osd_weight_total = 0; | |
5186 | map<int,float> osd_weight; | |
a8e16298 | 5187 | |
20effc67 TL |
5188 | if (max <= 0) { |
5189 | lderr(cct) << __func__ << " abort due to max <= 0" << dendl; | |
5190 | return 0; | |
a8e16298 | 5191 | } |
20effc67 TL |
5192 | |
5193 | osd_weight_total = build_pool_pgs_info(cct, only_pools, tmp_osd_map, | |
5194 | total_pgs, pgs_by_osd, osd_weight); | |
a8e16298 TL |
5195 | if (osd_weight_total == 0) { |
5196 | lderr(cct) << __func__ << " abort due to osd_weight_total == 0" << dendl; | |
5197 | return 0; | |
5198 | } | |
20effc67 | 5199 | |
a8e16298 TL |
5200 | float pgs_per_weight = total_pgs / osd_weight_total; |
5201 | ldout(cct, 10) << " osd_weight_total " << osd_weight_total << dendl; | |
5202 | ldout(cct, 10) << " pgs_per_weight " << pgs_per_weight << dendl; | |
7c673cae | 5203 | |
a8e16298 TL |
5204 | float stddev = 0; |
5205 | map<int,float> osd_deviation; // osd, deviation(pgs) | |
5206 | multimap<float,int> deviation_osd; // deviation(pgs), osd | |
20effc67 TL |
5207 | float cur_max_deviation = calc_deviations(cct, pgs_by_osd, osd_weight, pgs_per_weight, |
5208 | osd_deviation, deviation_osd, stddev); | |
5209 | ||
92f5a8d4 TL |
5210 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
5211 | if (cur_max_deviation <= max_deviation) { | |
a8e16298 TL |
5212 | ldout(cct, 10) << __func__ << " distribution is almost perfect" |
5213 | << dendl; | |
5214 | return 0; | |
5215 | } | |
20effc67 | 5216 | |
a8e16298 TL |
5217 | bool skip_overfull = false; |
5218 | auto aggressive = | |
11fdf7f2 | 5219 | cct->_conf.get_val<bool>("osd_calc_pg_upmaps_aggressively"); |
1e59de90 TL |
5220 | auto fast_aggressive = aggressive && |
5221 | cct->_conf.get_val<bool>("osd_calc_pg_upmaps_aggressively_fast"); | |
a8e16298 | 5222 | auto local_fallback_retries = |
11fdf7f2 | 5223 | cct->_conf.get_val<uint64_t>("osd_calc_pg_upmaps_local_fallback_retries"); |
20effc67 | 5224 | |
a8e16298 | 5225 | while (max--) { |
92f5a8d4 | 5226 | ldout(cct, 30) << "Top of loop #" << max+1 << dendl; |
a8e16298 TL |
5227 | // build overfull and underfull |
5228 | set<int> overfull; | |
92f5a8d4 TL |
5229 | set<int> more_overfull; |
5230 | bool using_more_overfull = false; | |
a8e16298 | 5231 | vector<int> underfull; |
92f5a8d4 | 5232 | vector<int> more_underfull; |
20effc67 TL |
5233 | fill_overfull_underfull(cct, deviation_osd, max_deviation, |
5234 | overfull, more_overfull, | |
5235 | underfull, more_underfull); | |
7c673cae | 5236 | |
92f5a8d4 TL |
5237 | if (underfull.empty() && overfull.empty()) { |
5238 | ldout(cct, 20) << __func__ << " failed to build overfull and underfull" << dendl; | |
7c673cae | 5239 | break; |
a8e16298 | 5240 | } |
92f5a8d4 TL |
5241 | if (overfull.empty() && !underfull.empty()) { |
5242 | ldout(cct, 20) << __func__ << " Using more_overfull since we still have underfull" << dendl; | |
5243 | overfull = more_overfull; | |
5244 | using_more_overfull = true; | |
5245 | } | |
7c673cae | 5246 | |
a8e16298 TL |
5247 | ldout(cct, 10) << " overfull " << overfull |
5248 | << " underfull " << underfull | |
5249 | << dendl; | |
5250 | set<pg_t> to_skip; | |
5251 | uint64_t local_fallback_retried = 0; | |
5252 | ||
1e59de90 TL |
5253 | // Used to prevent some of the unsuccessful loop iterations (save runtime) |
5254 | // If we can't find a change per OSD we skip further iterations for this OSD | |
5255 | uint n_changes = 0, prev_n_changes = 0; | |
5256 | set<int> osd_to_skip; | |
5257 | ||
a8e16298 TL |
5258 | retry: |
5259 | ||
5260 | set<pg_t> to_unmap; | |
5261 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>> to_upmap; | |
5262 | auto temp_pgs_by_osd = pgs_by_osd; | |
5263 | // always start with fullest, break if we find any changes to make | |
7c673cae | 5264 | for (auto p = deviation_osd.rbegin(); p != deviation_osd.rend(); ++p) { |
92f5a8d4 | 5265 | if (skip_overfull && !underfull.empty()) { |
a8e16298 TL |
5266 | ldout(cct, 10) << " skipping overfull " << dendl; |
5267 | break; // fall through to check underfull | |
5268 | } | |
7c673cae | 5269 | int osd = p->second; |
31f18b77 | 5270 | float deviation = p->first; |
1e59de90 TL |
5271 | if (fast_aggressive && osd_to_skip.count(osd)) { |
5272 | ldout(cct, 20) << " Fast aggressive mode: skipping osd " << osd | |
5273 | << " osd_to_skip size = " << osd_to_skip.size() << dendl; | |
5274 | continue; | |
5275 | } | |
5276 | ||
9f95a23c TL |
5277 | if (deviation < 0) { |
5278 | ldout(cct, 10) << " hitting underfull osds now" | |
5279 | << " when trying to remap overfull osds" | |
5280 | << dendl; | |
5281 | break; | |
5282 | } | |
7c673cae | 5283 | float target = osd_weight[osd] * pgs_per_weight; |
92f5a8d4 TL |
5284 | ldout(cct, 10) << " Overfull search osd." << osd |
5285 | << " target " << target | |
5286 | << " deviation " << deviation | |
5287 | << dendl; | |
a8e16298 | 5288 | ceph_assert(target > 0); |
92f5a8d4 | 5289 | if (!using_more_overfull && deviation <= max_deviation) { |
7c673cae | 5290 | ldout(cct, 10) << " osd." << osd |
a8e16298 TL |
5291 | << " target " << target |
5292 | << " deviation " << deviation | |
92f5a8d4 | 5293 | << " < max deviation " << max_deviation |
a8e16298 | 5294 | << dendl; |
7c673cae FG |
5295 | break; |
5296 | } | |
7c673cae | 5297 | |
a8e16298 TL |
5298 | vector<pg_t> pgs; |
5299 | pgs.reserve(pgs_by_osd[osd].size()); | |
5300 | for (auto& pg : pgs_by_osd[osd]) { | |
5301 | if (to_skip.count(pg)) | |
5302 | continue; | |
5303 | pgs.push_back(pg); | |
5304 | } | |
5305 | if (aggressive) { | |
5306 | // shuffle PG list so they all get equal (in)attention | |
20effc67 | 5307 | std::shuffle(pgs.begin(), pgs.end(), get_random_engine(cct, p_seed)); |
a8e16298 | 5308 | } |
7c673cae | 5309 | // look for remaps we can un-remap |
20effc67 TL |
5310 | if (try_drop_remap_overfull(cct, pgs, tmp_osd_map, osd, |
5311 | temp_pgs_by_osd, to_unmap, to_upmap)) | |
5312 | goto test_change; | |
7c673cae | 5313 | |
a8e16298 | 5314 | // try upmap |
7c673cae | 5315 | for (auto pg : pgs) { |
20effc67 TL |
5316 | auto temp_it = tmp_osd_map.pg_upmap.find(pg); |
5317 | if (temp_it != tmp_osd_map.pg_upmap.end()) { | |
a8e16298 TL |
5318 | // leave pg_upmap alone |
5319 | // it must be specified by admin since balancer does not | |
5320 | // support pg_upmap yet | |
5321 | ldout(cct, 10) << " " << pg << " already has pg_upmap " | |
5322 | << temp_it->second << ", skipping" | |
5323 | << dendl; | |
7c673cae FG |
5324 | continue; |
5325 | } | |
20effc67 | 5326 | auto pg_pool_size = tmp_osd_map.get_pg_pool_size(pg); |
a8e16298 TL |
5327 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; |
5328 | set<int> existing; | |
20effc67 TL |
5329 | auto it = tmp_osd_map.pg_upmap_items.find(pg); |
5330 | if (it != tmp_osd_map.pg_upmap_items.end()) { | |
5331 | auto& um_items = it->second; | |
5332 | if (um_items.size() >= (size_t)pg_pool_size) { | |
5333 | ldout(cct, 10) << " " << pg << " already has full-size pg_upmap_items " | |
5334 | << um_items << ", skipping" | |
5335 | << dendl; | |
5336 | continue; | |
5337 | } else { | |
5338 | ldout(cct, 10) << " " << pg << " already has pg_upmap_items " | |
5339 | << um_items | |
5340 | << dendl; | |
5341 | new_upmap_items = um_items; | |
5342 | // build existing too (for dedup) | |
5343 | for (auto [um_from, um_to] : um_items) { | |
5344 | existing.insert(um_from); | |
5345 | existing.insert(um_to); | |
5346 | } | |
5347 | } | |
a8e16298 TL |
5348 | // fall through |
5349 | // to see if we can append more remapping pairs | |
20effc67 | 5350 | } |
a8e16298 | 5351 | ldout(cct, 10) << " trying " << pg << dendl; |
494da23a | 5352 | vector<int> raw, orig, out; |
20effc67 | 5353 | tmp_osd_map.pg_to_raw_upmap(pg, &raw, &orig); // including existing upmaps too |
92f5a8d4 | 5354 | if (!try_pg_upmap(cct, pg, overfull, underfull, more_underfull, &orig, &out)) { |
7c673cae FG |
5355 | continue; |
5356 | } | |
a8e16298 | 5357 | ldout(cct, 10) << " " << pg << " " << orig << " -> " << out << dendl; |
7c673cae FG |
5358 | if (orig.size() != out.size()) { |
5359 | continue; | |
5360 | } | |
a8e16298 | 5361 | ceph_assert(orig != out); |
20effc67 | 5362 | int pos = find_best_remap(cct, orig, out, existing, osd_deviation); |
92f5a8d4 | 5363 | if (pos != -1) { |
a8e16298 TL |
5364 | // append new remapping pairs slowly |
5365 | // This way we can make sure that each tiny change will | |
5366 | // definitely make distribution of PGs converging to | |
5367 | // the perfect status. | |
20effc67 TL |
5368 | add_remap_pair(cct, orig[pos], out[pos], pg, (size_t)pg_pool_size, |
5369 | osd, existing, temp_pgs_by_osd, | |
5370 | new_upmap_items, to_upmap); | |
a8e16298 | 5371 | goto test_change; |
7c673cae | 5372 | } |
a8e16298 | 5373 | } |
1e59de90 TL |
5374 | if (fast_aggressive) { |
5375 | if (prev_n_changes == n_changes) { // no changes for prev OSD | |
5376 | osd_to_skip.insert(osd); | |
5377 | } | |
5378 | else { | |
5379 | prev_n_changes = n_changes; | |
5380 | } | |
5381 | } | |
5382 | ||
a8e16298 | 5383 | } |
7c673cae | 5384 | |
a8e16298 TL |
5385 | ceph_assert(!(to_unmap.size() || to_upmap.size())); |
5386 | ldout(cct, 10) << " failed to find any changes for overfull osds" | |
5387 | << dendl; | |
20effc67 TL |
5388 | for (auto& [deviation, osd] : deviation_osd) { |
5389 | if (std::find(underfull.begin(), underfull.end(), osd) == | |
a8e16298 TL |
5390 | underfull.end()) |
5391 | break; | |
a8e16298 TL |
5392 | float target = osd_weight[osd] * pgs_per_weight; |
5393 | ceph_assert(target > 0); | |
92f5a8d4 TL |
5394 | if (fabsf(deviation) < max_deviation) { |
5395 | // respect max_deviation too | |
a8e16298 TL |
5396 | ldout(cct, 10) << " osd." << osd |
5397 | << " target " << target | |
5398 | << " deviation " << deviation | |
92f5a8d4 TL |
5399 | << " -> absolute " << fabsf(deviation) |
5400 | << " < max " << max_deviation | |
a8e16298 TL |
5401 | << dendl; |
5402 | break; | |
5403 | } | |
5404 | // look for remaps we can un-remap | |
20effc67 TL |
5405 | candidates_t candidates = build_candidates(cct, tmp_osd_map, to_skip, |
5406 | only_pools, aggressive, p_seed); | |
5407 | if (try_drop_remap_underfull(cct, candidates, osd, temp_pgs_by_osd, | |
5408 | to_unmap, to_upmap)) { | |
5409 | goto test_change; | |
a8e16298 | 5410 | } |
7c673cae | 5411 | } |
a8e16298 TL |
5412 | |
5413 | ceph_assert(!(to_unmap.size() || to_upmap.size())); | |
5414 | ldout(cct, 10) << " failed to find any changes for underfull osds" | |
5415 | << dendl; | |
5416 | if (!aggressive) { | |
5417 | ldout(cct, 10) << " break due to aggressive mode not enabled" << dendl; | |
5418 | break; | |
5419 | } else if (!skip_overfull) { | |
5420 | // safe to quit because below here we know | |
5421 | // we've done checking both overfull and underfull osds.. | |
5422 | ldout(cct, 10) << " break due to not being able to find any" | |
5423 | << " further optimizations" | |
5424 | << dendl; | |
7c673cae FG |
5425 | break; |
5426 | } | |
a8e16298 TL |
5427 | // restart with fullest and do exhaustive searching |
5428 | skip_overfull = false; | |
5429 | continue; | |
5430 | ||
5431 | test_change: | |
5432 | ||
5433 | // test change, apply if change is good | |
5434 | ceph_assert(to_unmap.size() || to_upmap.size()); | |
5435 | float new_stddev = 0; | |
5436 | map<int,float> temp_osd_deviation; | |
5437 | multimap<float,int> temp_deviation_osd; | |
20effc67 TL |
5438 | float cur_max_deviation = calc_deviations(cct, temp_pgs_by_osd, osd_weight, |
5439 | pgs_per_weight, temp_osd_deviation, | |
5440 | temp_deviation_osd, new_stddev); | |
a8e16298 TL |
5441 | ldout(cct, 10) << " stddev " << stddev << " -> " << new_stddev << dendl; |
5442 | if (new_stddev >= stddev) { | |
5443 | if (!aggressive) { | |
5444 | ldout(cct, 10) << " break because stddev is not decreasing" | |
5445 | << " and aggressive mode is not enabled" | |
5446 | << dendl; | |
5447 | break; | |
5448 | } | |
5449 | local_fallback_retried++; | |
5450 | if (local_fallback_retried >= local_fallback_retries) { | |
5451 | // does not make progress | |
5452 | // flip *skip_overfull* so both overfull and underfull | |
5453 | // get equal (in)attention | |
5454 | skip_overfull = !skip_overfull; | |
5455 | ldout(cct, 10) << " hit local_fallback_retries " | |
5456 | << local_fallback_retries | |
5457 | << dendl; | |
5458 | continue; | |
5459 | } | |
5460 | for (auto& i : to_unmap) | |
5461 | to_skip.insert(i); | |
5462 | for (auto& i : to_upmap) | |
5463 | to_skip.insert(i.first); | |
5464 | ldout(cct, 20) << " local_fallback_retried " << local_fallback_retried | |
5465 | << " to_skip " << to_skip | |
5466 | << dendl; | |
5467 | goto retry; | |
5468 | } | |
5469 | ||
5470 | // ready to go | |
5471 | ceph_assert(new_stddev < stddev); | |
5472 | stddev = new_stddev; | |
5473 | pgs_by_osd = temp_pgs_by_osd; | |
5474 | osd_deviation = temp_osd_deviation; | |
5475 | deviation_osd = temp_deviation_osd; | |
1e59de90 TL |
5476 | n_changes++; |
5477 | ||
20effc67 TL |
5478 | |
5479 | num_changed += pack_upmap_results(cct, to_unmap, to_upmap, tmp_osd_map, pending_inc); | |
5480 | ||
92f5a8d4 TL |
5481 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
5482 | if (cur_max_deviation <= max_deviation) { | |
5483 | ldout(cct, 10) << __func__ << " Optimization plan is almost perfect" | |
5484 | << dendl; | |
5485 | break; | |
5486 | } | |
7c673cae | 5487 | } |
a8e16298 | 5488 | ldout(cct, 10) << " num_changed = " << num_changed << dendl; |
7c673cae FG |
5489 | return num_changed; |
5490 | } | |
31f18b77 | 5491 | |
1e59de90 TL |
5492 | map<uint64_t,set<pg_t>> OSDMap::get_pgs_by_osd( |
5493 | CephContext *cct, | |
5494 | int64_t pid, | |
5495 | map<uint64_t, set<pg_t>> *p_primaries_by_osd, | |
5496 | map<uint64_t, set<pg_t>> *p_acting_primaries_by_osd) const | |
5497 | { | |
5498 | // Set up the OSDMap | |
5499 | OSDMap tmp_osd_map; | |
5500 | tmp_osd_map.deepish_copy_from(*this); | |
5501 | ||
5502 | // Get the pool from the provided pool id | |
5503 | const pg_pool_t* pool = get_pg_pool(pid); | |
5504 | ||
5505 | // build array of pgs from the pool | |
5506 | map<uint64_t,set<pg_t>> pgs_by_osd; | |
5507 | for (unsigned ps = 0; ps < pool->get_pg_num(); ++ps) { | |
5508 | pg_t pg(ps, pid); | |
5509 | vector<int> up; | |
5510 | int primary; | |
5511 | int acting_prim; | |
5512 | tmp_osd_map.pg_to_up_acting_osds(pg, &up, &primary, nullptr, &acting_prim); | |
5513 | if (cct != nullptr) | |
5514 | ldout(cct, 20) << __func__ << " " << pg | |
5515 | << " up " << up | |
5516 | << " primary " << primary | |
5517 | << " acting_primary " << acting_prim | |
5518 | << dendl; | |
5519 | ||
5520 | if (!up.empty()) { // up can be empty is test generated files | |
5521 | // in this case, we return empty result | |
5522 | for (auto osd : up) { | |
5523 | if (osd != CRUSH_ITEM_NONE) | |
5524 | pgs_by_osd[osd].insert(pg); | |
5525 | } | |
5526 | if (p_primaries_by_osd != nullptr) { | |
5527 | if (primary != CRUSH_ITEM_NONE) | |
5528 | (*p_primaries_by_osd)[primary].insert(pg); | |
5529 | } | |
5530 | if (p_acting_primaries_by_osd != nullptr) { | |
5531 | if (acting_prim != CRUSH_ITEM_NONE) | |
5532 | (*p_acting_primaries_by_osd)[acting_prim].insert(pg); | |
5533 | } | |
5534 | } | |
5535 | } | |
5536 | return pgs_by_osd; | |
5537 | } | |
5538 | ||
5539 | float OSDMap::get_osds_weight( | |
5540 | CephContext *cct, | |
5541 | const OSDMap& tmp_osd_map, | |
5542 | int64_t pid, | |
5543 | map<int,float>& osds_weight) const | |
5544 | { | |
5545 | map<int,float> pmap; | |
5546 | ceph_assert(pools.count(pid)); | |
5547 | int ruleno = pools.at(pid).get_crush_rule(); | |
5548 | tmp_osd_map.crush->get_rule_weight_osd_map(ruleno, &pmap); | |
5549 | ldout(cct,20) << __func__ << " pool " << pid | |
5550 | << " ruleno " << ruleno | |
5551 | << " weight-map " << pmap | |
5552 | << dendl; | |
5553 | float osds_weight_total = 0; | |
5554 | for (auto [oid, oweight] : pmap) { | |
5555 | auto adjusted_weight = tmp_osd_map.get_weightf(oid) * oweight; | |
5556 | if (adjusted_weight != 0) { | |
5557 | osds_weight[oid] += adjusted_weight; | |
5558 | osds_weight_total += adjusted_weight; | |
5559 | } | |
5560 | } | |
5561 | return osds_weight_total; | |
5562 | } | |
5563 | ||
20effc67 TL |
5564 | float OSDMap::build_pool_pgs_info ( |
5565 | CephContext *cct, | |
5566 | const std::set<int64_t>& only_pools, ///< [optional] restrict to pool | |
5567 | const OSDMap& tmp_osd_map, | |
5568 | int& total_pgs, | |
5569 | map<int,set<pg_t>>& pgs_by_osd, | |
1e59de90 | 5570 | map<int,float>& osds_weight) |
20effc67 TL |
5571 | { |
5572 | // | |
5573 | // This function builds some data structures that are used by calc_pg_upmaps. | |
5574 | // Specifically it builds pgs_by_osd and osd_weight maps, updates total_pgs | |
5575 | // and returns the osd_weight_total | |
5576 | // | |
1e59de90 | 5577 | float osds_weight_total = 0.0; |
20effc67 TL |
5578 | for (auto& [pid, pdata] : pools) { |
5579 | if (!only_pools.empty() && !only_pools.count(pid)) | |
5580 | continue; | |
5581 | for (unsigned ps = 0; ps < pdata.get_pg_num(); ++ps) { | |
5582 | pg_t pg(ps, pid); | |
5583 | vector<int> up; | |
5584 | tmp_osd_map.pg_to_up_acting_osds(pg, &up, nullptr, nullptr, nullptr); | |
5585 | ldout(cct, 20) << __func__ << " " << pg << " up " << up << dendl; | |
5586 | for (auto osd : up) { | |
5587 | if (osd != CRUSH_ITEM_NONE) | |
5588 | pgs_by_osd[osd].insert(pg); | |
5589 | } | |
5590 | } | |
5591 | total_pgs += pdata.get_size() * pdata.get_pg_num(); | |
5592 | ||
1e59de90 | 5593 | osds_weight_total = get_osds_weight(cct, tmp_osd_map, pid, osds_weight); |
20effc67 | 5594 | } |
1e59de90 | 5595 | for (auto& [oid, oweight] : osds_weight) { |
20effc67 TL |
5596 | int pgs = 0; |
5597 | auto p = pgs_by_osd.find(oid); | |
5598 | if (p != pgs_by_osd.end()) | |
5599 | pgs = p->second.size(); | |
5600 | else | |
5601 | pgs_by_osd.emplace(oid, set<pg_t>()); | |
5602 | ldout(cct, 20) << " osd." << oid << " weight " << oweight | |
5603 | << " pgs " << pgs << dendl; | |
5604 | } | |
1e59de90 | 5605 | return osds_weight_total; |
20effc67 TL |
5606 | |
5607 | } // return total weight of all OSDs | |
5608 | ||
5609 | float OSDMap::calc_deviations ( | |
5610 | CephContext *cct, | |
5611 | const map<int,set<pg_t>>& pgs_by_osd, | |
5612 | const map<int,float>& osd_weight, | |
5613 | float pgs_per_weight, | |
5614 | map<int,float>& osd_deviation, | |
5615 | multimap<float,int>& deviation_osd, | |
5616 | float& stddev) // return current max deviation | |
5617 | { | |
5618 | // | |
5619 | // This function calculates the 2 maps osd_deviation and deviation_osd which | |
5620 | // hold the deviation between the current number of PGs which map to an OSD | |
5621 | // and the optimal number. Ot also calculates the stddev of the deviations and | |
5622 | // returns the current max deviation. | |
5623 | // NOTE - the calculation is not exactly stddev it is actually sttdev^2 but as | |
5624 | // long as it is monotonic with stddev (and it is), it is sufficient for | |
5625 | // the balancer code. | |
5626 | // | |
5627 | float cur_max_deviation = 0.0; | |
5628 | stddev = 0.0; | |
5629 | for (auto& [oid, opgs] : pgs_by_osd) { | |
5630 | // make sure osd is still there (belongs to this crush-tree) | |
5631 | ceph_assert(osd_weight.count(oid)); | |
5632 | float target = osd_weight.at(oid) * pgs_per_weight; | |
5633 | float deviation = (float)opgs.size() - target; | |
5634 | ldout(cct, 20) << " osd." << oid | |
5635 | << "\tpgs " << opgs.size() | |
5636 | << "\ttarget " << target | |
5637 | << "\tdeviation " << deviation | |
5638 | << dendl; | |
5639 | osd_deviation[oid] = deviation; | |
5640 | deviation_osd.insert(make_pair(deviation, oid)); | |
5641 | stddev += deviation * deviation; | |
5642 | if (fabsf(deviation) > cur_max_deviation) | |
5643 | cur_max_deviation = fabsf(deviation); | |
5644 | } | |
5645 | return cur_max_deviation; | |
5646 | } | |
5647 | ||
5648 | void OSDMap::fill_overfull_underfull ( | |
5649 | CephContext *cct, | |
5650 | const std::multimap<float,int>& deviation_osd, | |
5651 | int max_deviation, | |
5652 | std::set<int>& overfull, | |
5653 | std::set<int>& more_overfull, | |
5654 | std::vector<int>& underfull, | |
5655 | std::vector<int>& more_underfull) | |
5656 | { | |
5657 | // | |
5658 | // This function just fills the overfull and underfull data structures for the | |
5659 | // use of calc_pg_upmaps | |
5660 | // | |
5661 | for (auto i = deviation_osd.rbegin(); i != deviation_osd.rend(); i++) { | |
5662 | auto& odev = i->first; | |
5663 | auto& oid = i->second; | |
5664 | ldout(cct, 30) << " check " << odev << " <= " << max_deviation << dendl; | |
5665 | if (odev <= 0) | |
5666 | break; | |
5667 | if (odev > max_deviation) { | |
5668 | ldout(cct, 30) << " add overfull osd." << oid << dendl; | |
5669 | overfull.insert(oid); | |
5670 | } else { | |
5671 | more_overfull.insert(oid); | |
5672 | } | |
5673 | } | |
5674 | ||
5675 | for (auto i = deviation_osd.begin(); i != deviation_osd.end(); i++) { | |
5676 | auto& odev = i->first; | |
5677 | auto& oid = i->second; | |
5678 | ldout(cct, 30) << " check " << odev << " >= " << -(int)max_deviation << dendl; | |
5679 | if (odev >= 0) | |
5680 | break; | |
5681 | if (odev < -(int)max_deviation) { | |
5682 | ldout(cct, 30) << " add underfull osd." << oid << dendl; | |
5683 | underfull.push_back(oid); | |
5684 | } else { | |
5685 | more_underfull.push_back(oid); | |
5686 | } | |
5687 | } | |
5688 | } | |
5689 | ||
5690 | int OSDMap::pack_upmap_results( | |
5691 | CephContext *cct, | |
5692 | const std::set<pg_t>& to_unmap, | |
5693 | const std::map<pg_t, mempool::osdmap::vector<std::pair<int, int>>>& to_upmap, | |
5694 | OSDMap& tmp_osd_map, | |
5695 | OSDMap::Incremental *pending_inc) | |
5696 | { | |
5697 | // | |
5698 | // This function takes the input from the local variables to_unmap and to_upmap | |
5699 | // and updates tmp_osd_map (so that another iteration can run) and pending_inc | |
5700 | // (so that the results are visible outside calc_pg_upmaps) | |
5701 | // | |
5702 | int num_changed = 0; | |
5703 | for (auto& i : to_unmap) { | |
5704 | ldout(cct, 10) << " unmap pg " << i << dendl; | |
5705 | ceph_assert(tmp_osd_map.pg_upmap_items.count(i)); | |
5706 | tmp_osd_map.pg_upmap_items.erase(i); | |
5707 | pending_inc->old_pg_upmap_items.insert(i); | |
5708 | ++num_changed; | |
5709 | } | |
5710 | for (auto& [pg, um_items] : to_upmap) { | |
5711 | ldout(cct, 10) << " upmap pg " << pg | |
5712 | << " new pg_upmap_items " << um_items | |
5713 | << dendl; | |
5714 | tmp_osd_map.pg_upmap_items[pg] = um_items; | |
5715 | pending_inc->new_pg_upmap_items[pg] = um_items; | |
5716 | ++num_changed; | |
5717 | } | |
5718 | ||
5719 | return num_changed; | |
5720 | } | |
5721 | ||
5722 | std::default_random_engine OSDMap::get_random_engine( | |
5723 | CephContext *cct, | |
5724 | std::random_device::result_type *p_seed) | |
5725 | { | |
5726 | // | |
5727 | // This function creates a random_engine to be used for shuffling. | |
5728 | // When p_seed == nullptr it generates random engine with a seed from /dev/random | |
5729 | // when p_seed is not null, it uses (*p_seed + seed_set) as the seed and | |
5730 | // increments seed_set. This is used in order to craete regression test without | |
5731 | // random effect on the results. | |
5732 | // | |
5733 | static std::random_device::result_type seed_set = 0; | |
5734 | std::random_device::result_type seed; | |
5735 | if (p_seed == nullptr) { | |
5736 | std::random_device rd; | |
5737 | seed = rd(); | |
5738 | } | |
5739 | else { | |
5740 | seed = *p_seed + seed_set; | |
5741 | ldout(cct, 30) << " Starting random engine with seed " | |
5742 | << seed << dendl; | |
5743 | seed_set++; | |
5744 | } | |
5745 | return std::default_random_engine{seed}; | |
5746 | } | |
5747 | ||
5748 | bool OSDMap::try_drop_remap_overfull( | |
5749 | CephContext *cct, | |
5750 | const std::vector<pg_t>& pgs, | |
5751 | const OSDMap& tmp_osd_map, | |
5752 | int osd, | |
5753 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5754 | set<pg_t>& to_unmap, | |
5755 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5756 | { | |
5757 | // | |
5758 | // This function tries to drop existimg upmap items which map data to overfull | |
5759 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5760 | // if it found an item that can be dropped, false if not. | |
5761 | // | |
5762 | for (auto pg : pgs) { | |
5763 | auto p = tmp_osd_map.pg_upmap_items.find(pg); | |
5764 | if (p == tmp_osd_map.pg_upmap_items.end()) | |
5765 | continue; | |
5766 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5767 | auto& pg_upmap_items = p->second; | |
5768 | for (auto um_pair : pg_upmap_items) { | |
5769 | auto& um_from = um_pair.first; | |
5770 | auto& um_to = um_pair.second; | |
5771 | if (um_to == osd) { | |
5772 | ldout(cct, 10) << " will try dropping existing" | |
5773 | << " remapping pair " | |
5774 | << um_from << " -> " << um_to | |
5775 | << " which remapped " << pg | |
5776 | << " into overfull osd." << osd | |
5777 | << dendl; | |
5778 | temp_pgs_by_osd[um_to].erase(pg); | |
5779 | temp_pgs_by_osd[um_from].insert(pg); | |
5780 | } else { | |
5781 | new_upmap_items.push_back(um_pair); | |
5782 | } | |
5783 | } | |
5784 | if (new_upmap_items.empty()) { | |
5785 | // drop whole item | |
5786 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5787 | << " remapped " << pg << " into overfull osd." << osd | |
5788 | << ", will try cancelling it entirely" | |
5789 | << dendl; | |
5790 | to_unmap.insert(pg); | |
5791 | return true; | |
5792 | } else if (new_upmap_items.size() != pg_upmap_items.size()) { | |
5793 | // drop single remapping pair, updating | |
5794 | ceph_assert(new_upmap_items.size() < pg_upmap_items.size()); | |
5795 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5796 | << " remapped " << pg << " into overfull osd." << osd | |
5797 | << ", new_pg_upmap_items now " << new_upmap_items | |
5798 | << dendl; | |
5799 | to_upmap[pg] = new_upmap_items; | |
5800 | return true; | |
5801 | } | |
5802 | } | |
5803 | return false; | |
5804 | } | |
5805 | ||
5806 | bool OSDMap::try_drop_remap_underfull( | |
5807 | CephContext *cct, | |
5808 | const candidates_t& candidates, | |
5809 | int osd, | |
5810 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5811 | set<pg_t>& to_unmap, | |
5812 | map<pg_t, mempool::osdmap::vector<std::pair<int32_t,int32_t>>>& to_upmap) | |
5813 | { | |
5814 | // | |
5815 | // This function tries to drop existimg upmap items which map data from underfull | |
5816 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5817 | // if it found an item that can be dropped, false if not. | |
5818 | // | |
5819 | for (auto& [pg, um_pairs] : candidates) { | |
5820 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5821 | for (auto& ump : um_pairs) { | |
5822 | auto& um_from = ump.first; | |
5823 | auto& um_to = ump.second; | |
5824 | if (um_from == osd) { | |
5825 | ldout(cct, 10) << " will try dropping existing" | |
5826 | << " remapping pair " | |
5827 | << um_from << " -> " << um_to | |
5828 | << " which remapped " << pg | |
5829 | << " out from underfull osd." << osd | |
5830 | << dendl; | |
5831 | temp_pgs_by_osd[um_to].erase(pg); | |
5832 | temp_pgs_by_osd[um_from].insert(pg); | |
5833 | } else { | |
5834 | new_upmap_items.push_back(ump); | |
5835 | } | |
5836 | } | |
5837 | if (new_upmap_items.empty()) { | |
5838 | // drop whole item | |
5839 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5840 | << " remapped " << pg | |
5841 | << " out from underfull osd." << osd | |
5842 | << ", will try cancelling it entirely" | |
5843 | << dendl; | |
5844 | to_unmap.insert(pg); | |
5845 | return true; | |
5846 | } else if (new_upmap_items.size() != um_pairs.size()) { | |
5847 | // drop single remapping pair, updating | |
5848 | ceph_assert(new_upmap_items.size() < um_pairs.size()); | |
5849 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5850 | << " remapped " << pg | |
5851 | << " out from underfull osd." << osd | |
5852 | << ", new_pg_upmap_items now " << new_upmap_items | |
5853 | << dendl; | |
5854 | to_upmap[pg] = new_upmap_items; | |
5855 | return true; | |
5856 | } | |
5857 | } | |
5858 | return false; | |
5859 | } | |
5860 | ||
5861 | void OSDMap::add_remap_pair( | |
5862 | CephContext *cct, | |
5863 | int orig, | |
5864 | int out, | |
5865 | pg_t pg, | |
5866 | size_t pg_pool_size, | |
5867 | int osd, | |
5868 | set<int>& existing, | |
5869 | map<int,set<pg_t>>& temp_pgs_by_osd, | |
5870 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items, | |
5871 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5872 | { | |
5873 | // | |
5874 | // add a single remap pair (in pg <pg> remap osd from <orig> to <out>) to all | |
5875 | // the relevant data structures | |
5876 | // | |
5877 | ldout(cct, 10) << " will try adding new remapping pair " | |
5878 | << orig << " -> " << out << " for " << pg | |
5879 | << (orig != osd ? " NOT selected osd" : "") | |
5880 | << dendl; | |
5881 | existing.insert(orig); | |
5882 | existing.insert(out); | |
5883 | temp_pgs_by_osd[orig].erase(pg); | |
5884 | temp_pgs_by_osd[out].insert(pg); | |
5885 | ceph_assert(new_upmap_items.size() < pg_pool_size); | |
5886 | new_upmap_items.push_back(make_pair(orig, out)); | |
5887 | // append new remapping pairs slowly | |
5888 | // This way we can make sure that each tiny change will | |
5889 | // definitely make distribution of PGs converging to | |
5890 | // the perfect status. | |
5891 | to_upmap[pg] = new_upmap_items; | |
5892 | ||
5893 | } | |
5894 | ||
5895 | int OSDMap::find_best_remap ( | |
5896 | CephContext *cct, | |
5897 | const vector<int>& orig, | |
5898 | const vector<int>& out, | |
5899 | const set<int>& existing, | |
5900 | const map<int,float> osd_deviation) | |
5901 | { | |
5902 | // | |
5903 | // Find the best remap from the suggestions in orig and out - the best remap | |
5904 | // is the one which maps from the OSD with the largest deviatoion (from the | |
5905 | // OSDs which are part of orig) | |
5906 | // | |
5907 | int best_pos = -1; | |
5908 | float max_dev = 0; | |
5909 | for (unsigned i = 0; i < out.size(); ++i) { | |
5910 | if (orig[i] == out[i]) | |
5911 | continue; // skip invalid remappings | |
5912 | if (existing.count(orig[i]) || existing.count(out[i])) | |
5913 | continue; // we want new remappings only! | |
5914 | if (osd_deviation.at(orig[i]) > max_dev) { | |
5915 | max_dev = osd_deviation.at(orig[i]); | |
5916 | best_pos = i; | |
5917 | ldout(cct, 30) << "Max osd." << orig[i] << " pos " << i << " dev " << osd_deviation.at(orig[i]) << dendl; | |
5918 | } | |
5919 | } | |
5920 | return best_pos; | |
5921 | } | |
5922 | ||
5923 | OSDMap::candidates_t OSDMap::build_candidates( | |
5924 | CephContext *cct, | |
5925 | const OSDMap& tmp_osd_map, | |
5926 | const set<pg_t> to_skip, | |
5927 | const set<int64_t>& only_pools, | |
5928 | bool aggressive, | |
5929 | std::random_device::result_type *p_seed) | |
5930 | { | |
5931 | // | |
5932 | // build the candidates data structure | |
5933 | // | |
5934 | candidates_t candidates; | |
5935 | candidates.reserve(tmp_osd_map.pg_upmap_items.size()); | |
5936 | for (auto& [pg, um_pair] : tmp_osd_map.pg_upmap_items) { | |
5937 | if (to_skip.count(pg)) | |
5938 | continue; | |
5939 | if (!only_pools.empty() && !only_pools.count(pg.pool())) | |
5940 | continue; | |
5941 | candidates.push_back(make_pair(pg, um_pair)); | |
5942 | } | |
5943 | if (aggressive) { | |
5944 | // shuffle candidates so they all get equal (in)attention | |
5945 | std::shuffle(candidates.begin(), candidates.end(), get_random_engine(cct, p_seed)); | |
5946 | } | |
5947 | return candidates; | |
5948 | } | |
5949 | ||
1e59de90 TL |
5950 | // return -1 if all PGs are OK, else the first PG which includes only zero PA OSDs |
5951 | int64_t OSDMap::has_zero_pa_pgs(CephContext *cct, int64_t pool_id) const | |
5952 | { | |
5953 | const pg_pool_t* pool = get_pg_pool(pool_id); | |
5954 | for (unsigned ps = 0; ps < pool->get_pg_num(); ++ps) { | |
5955 | pg_t pg(ps, pool_id); | |
5956 | vector<int> acting; | |
5957 | pg_to_up_acting_osds(pg, nullptr, nullptr, &acting, nullptr); | |
5958 | if (cct != nullptr) { | |
5959 | ldout(cct, 30) << __func__ << " " << pg << " acting " << acting << dendl; | |
5960 | } | |
5961 | bool pg_zero_pa = true; | |
5962 | for (auto osd : acting) { | |
5963 | if (get_primary_affinityf(osd) != 0) { | |
5964 | pg_zero_pa = false; | |
5965 | break; | |
5966 | } | |
5967 | } | |
5968 | if (pg_zero_pa) { | |
5969 | if (cct != nullptr) { | |
5970 | ldout(cct, 20) << __func__ << " " << pg << " - maps only to OSDs with primiary affinity 0" << dendl; | |
5971 | } | |
5972 | return (int64_t)ps; | |
5973 | } | |
5974 | } | |
5975 | return -1; | |
5976 | } | |
5977 | ||
5978 | void OSDMap::zero_rbi(read_balance_info_t &rbi) const { | |
5979 | rbi.pa_avg = 0.; | |
5980 | rbi.pa_weighted = 0.; | |
5981 | rbi.pa_weighted_avg = 0.; | |
5982 | rbi.raw_score = 0.; | |
5983 | rbi.optimal_score = 0.; | |
5984 | rbi.adjusted_score = 0.; | |
5985 | rbi.acting_raw_score = 0.; | |
5986 | rbi.acting_adj_score = 0.; | |
5987 | rbi.err_msg = ""; | |
5988 | } | |
5989 | ||
5990 | int OSDMap::set_rbi( | |
5991 | CephContext *cct, | |
5992 | read_balance_info_t &rbi, | |
5993 | int64_t pool_id, | |
5994 | float total_w_pa, | |
5995 | float pa_sum, | |
5996 | int num_osds, | |
5997 | int osd_pa_count, | |
5998 | float total_osd_weight, | |
5999 | uint max_prims_per_osd, | |
6000 | uint max_acting_prims_per_osd, | |
6001 | float avg_prims_per_osd, | |
6002 | bool prim_on_zero_pa, | |
6003 | bool acting_on_zero_pa, | |
6004 | float max_osd_score) const | |
6005 | { | |
6006 | // put all the ugly code here, so rest of code is nicer. | |
6007 | const pg_pool_t* pool = get_pg_pool(pool_id); | |
6008 | zero_rbi(rbi); | |
6009 | ||
6010 | if (total_w_pa / total_osd_weight < 1. / float(pool->get_size())) { | |
6011 | ldout(cct, 20) << __func__ << " pool " << pool_id << " average primary affinity is lower than" | |
6012 | << 1. / float(pool->get_size()) << dendl; | |
6013 | rbi.err_msg = fmt::format( | |
6014 | "pool {} average primary affinity is lower than {:.2f}, read balance score is not reliable", | |
6015 | pool_id, 1. / float(pool->get_size())); | |
6016 | return -EINVAL; | |
6017 | } | |
6018 | rbi.pa_weighted = total_w_pa; | |
6019 | ||
6020 | // weighted_prim_affinity_avg | |
6021 | rbi.pa_weighted_avg = rbi_round(rbi.pa_weighted / total_osd_weight); // in [0..1] | |
6022 | // p_rbi->pa_weighted / osd_pa_count; // in [0..1] | |
6023 | ||
6024 | rbi.raw_score = rbi_round((float)max_prims_per_osd / avg_prims_per_osd); // >=1 | |
6025 | if (acting_on_zero_pa) { | |
6026 | rbi.acting_raw_score = rbi_round(max_osd_score); | |
6027 | rbi.err_msg = fmt::format( | |
6028 | "pool {} has acting primaries on OSD(s) with primary affinity 0, read balance score is not accurate", | |
6029 | pool_id); | |
6030 | } else { | |
6031 | rbi.acting_raw_score = rbi_round((float)max_acting_prims_per_osd / avg_prims_per_osd); | |
6032 | } | |
6033 | ||
6034 | if (osd_pa_count != 0) { | |
6035 | // this implies that pa_sum > 0 | |
6036 | rbi.pa_avg = rbi_round(pa_sum / osd_pa_count); // in [0..1] | |
6037 | } else { | |
6038 | rbi.pa_avg = 0.; | |
6039 | } | |
6040 | ||
6041 | if (rbi.pa_avg != 0.) { | |
6042 | int64_t zpg; | |
6043 | if ((zpg = has_zero_pa_pgs(cct, pool_id)) >= 0) { | |
6044 | pg_t pg(zpg, pool_id); | |
6045 | std::stringstream ss; | |
6046 | ss << pg; | |
6047 | ldout(cct, 10) << __func__ << " pool " << pool_id << " has some PGs where all OSDs are with primary_affinity 0 (" << pg << ",...)" << dendl; | |
6048 | rbi.err_msg = fmt::format( | |
6049 | "pool {} has some PGs where all OSDs are with primary_affinity 0 (at least pg {}), read balance score may not be reliable", | |
6050 | pool_id, ss.str()); | |
6051 | return -EINVAL; | |
6052 | } | |
6053 | rbi.optimal_score = rbi_round(float(num_osds) / float(osd_pa_count)); // >= 1 | |
6054 | // adjust the score to the primary affinity setting (if prim affinity is set | |
6055 | // the raw score can't be 1 and the optimal (perfect) score is hifgher than 1) | |
6056 | // When total system primary affinity is too low (average < 1 / pool replica count) | |
6057 | // the score is negative in order to grab the user's attention. | |
6058 | rbi.adjusted_score = rbi_round(rbi.raw_score / rbi.optimal_score); // >= 1 if PA is not low | |
6059 | rbi.acting_adj_score = rbi_round(rbi.acting_raw_score / rbi.optimal_score); // >= 1 if PA is not low | |
6060 | ||
6061 | } else { | |
6062 | // We should never get here - this condition is checked before calling this function - this is just sanity check code. | |
6063 | rbi.err_msg = fmt::format( | |
6064 | "pool {} all OSDs have zero primary affinity, can't calculate a reliable read balance score", | |
6065 | pool_id); | |
6066 | return -EINVAL; | |
6067 | } | |
6068 | ||
6069 | return 0; | |
6070 | } | |
6071 | ||
6072 | int OSDMap::calc_read_balance_score(CephContext *cct, int64_t pool_id, | |
6073 | read_balance_info_t *p_rbi) const | |
6074 | { | |
6075 | //BUG: wrong score with one PG replica 3 and 4 OSDs | |
6076 | if (cct != nullptr) | |
6077 | ldout(cct,20) << __func__ << " pool " << get_pool_name(pool_id) << dendl; | |
6078 | ||
6079 | OSDMap tmp_osd_map; | |
6080 | tmp_osd_map.deepish_copy_from(*this); | |
6081 | if (p_rbi == nullptr) { | |
6082 | // The only case where error message is not set - this is not tested in the unit test. | |
6083 | if (cct != nullptr) | |
6084 | ldout(cct,30) << __func__ << " p_rbi is nullptr." << dendl; | |
6085 | return -EINVAL; | |
6086 | } | |
6087 | ||
6088 | if (tmp_osd_map.pools.count(pool_id) == 0) { | |
6089 | if (cct != nullptr) | |
6090 | ldout(cct,30) << __func__ << " pool " << pool_id << " not found." << dendl; | |
6091 | zero_rbi(*p_rbi); | |
6092 | p_rbi->err_msg = fmt::format("pool {} not found", pool_id); | |
6093 | return -ENOENT; | |
6094 | } | |
6095 | int rc = 0; | |
6096 | const pg_pool_t* pool = tmp_osd_map.get_pg_pool(pool_id); | |
6097 | auto num_pgs = pool->get_pg_num(); | |
6098 | ||
6099 | map<uint64_t,set<pg_t>> pgs_by_osd; | |
6100 | map<uint64_t,set<pg_t>> prim_pgs_by_osd; | |
6101 | map<uint64_t,set<pg_t>> acting_prims_by_osd; | |
6102 | ||
6103 | pgs_by_osd = tmp_osd_map.get_pgs_by_osd(cct, pool_id, &prim_pgs_by_osd, &acting_prims_by_osd); | |
6104 | ||
6105 | if (cct != nullptr) | |
6106 | ldout(cct,30) << __func__ << " Primaries for pool: " | |
6107 | << prim_pgs_by_osd << dendl; | |
6108 | ||
6109 | if (pgs_by_osd.empty()) { | |
6110 | //p_rbi->err_msg = fmt::format("pool {} has no PGs mapped to OSDs", pool_id); | |
6111 | return -EINVAL; | |
6112 | } | |
6113 | if (cct != nullptr) { | |
6114 | for (auto& [osd,pgs] : prim_pgs_by_osd) { | |
6115 | ldout(cct,20) << __func__ << " Pool " << pool_id << " OSD." << osd | |
6116 | << " has " << pgs.size() << " primary PGs, " | |
6117 | << acting_prims_by_osd[osd].size() << " acting primaries." | |
6118 | << dendl; | |
6119 | } | |
6120 | } | |
6121 | ||
6122 | auto num_osds = pgs_by_osd.size(); | |
6123 | ||
6124 | float avg_prims_per_osd = (float)num_pgs / (float)num_osds; | |
6125 | uint64_t max_prims_per_osd = 0; | |
6126 | uint64_t max_acting_prims_per_osd = 0; | |
6127 | float max_osd_score = 0.; | |
6128 | bool prim_on_zero_pa = false; | |
6129 | bool acting_on_zero_pa = false; | |
6130 | ||
6131 | float prim_affinity_sum = 0.; | |
6132 | float total_osd_weight = 0.; | |
6133 | float total_weighted_pa = 0.; | |
6134 | ||
6135 | map<int,float> osds_crush_weight; | |
6136 | // Set up the OSDMap | |
6137 | int ruleno = tmp_osd_map.pools.at(pool_id).get_crush_rule(); | |
6138 | tmp_osd_map.crush->get_rule_weight_osd_map(ruleno, &osds_crush_weight); | |
6139 | ||
6140 | if (cct != nullptr) { | |
6141 | ldout(cct,20) << __func__ << " pool " << pool_id | |
6142 | << " ruleno " << ruleno | |
6143 | << " weight-map " << osds_crush_weight | |
6144 | << dendl; | |
6145 | } | |
6146 | uint osd_pa_count = 0; | |
6147 | ||
6148 | for (auto [osd, oweight] : osds_crush_weight) { // loop over all OSDs | |
6149 | total_osd_weight += oweight; | |
6150 | float osd_pa = tmp_osd_map.get_primary_affinityf(osd); | |
6151 | total_weighted_pa += oweight * osd_pa; | |
6152 | if (osd_pa != 0.) { | |
6153 | osd_pa_count++; | |
6154 | } | |
6155 | if (prim_pgs_by_osd.count(osd)) { | |
6156 | auto n_prims = prim_pgs_by_osd.at(osd).size(); | |
6157 | max_prims_per_osd = std::max(max_prims_per_osd, n_prims); | |
6158 | if (osd_pa == 0.) { | |
6159 | prim_on_zero_pa = true; | |
6160 | } | |
6161 | } | |
6162 | if (acting_prims_by_osd.count(osd)) { | |
6163 | auto n_aprims = acting_prims_by_osd.at(osd).size(); | |
6164 | max_acting_prims_per_osd = std::max(max_acting_prims_per_osd, n_aprims); | |
6165 | if (osd_pa != 0.) { | |
6166 | max_osd_score = std::max(max_osd_score, float(n_aprims) / osd_pa); | |
6167 | } | |
6168 | else { | |
6169 | acting_on_zero_pa = true; | |
6170 | } | |
6171 | } | |
6172 | ||
6173 | prim_affinity_sum += osd_pa; | |
6174 | if (cct != nullptr) { | |
6175 | auto np = prim_pgs_by_osd.count(osd) ? prim_pgs_by_osd.at(osd).size() : 0; | |
6176 | auto nap = acting_prims_by_osd.count(osd) ? acting_prims_by_osd.at(osd).size() : 0; | |
6177 | auto wt = osds_crush_weight.count(osd) ? osds_crush_weight.at(osd) : 0.; | |
6178 | ldout(cct,30) << __func__ << " OSD." << osd << " info: " | |
6179 | << " num_primaries " << np | |
6180 | << " num_acting_prims " << nap | |
6181 | << " prim_affinity " << tmp_osd_map.get_primary_affinityf(osd) | |
6182 | << " weight " << wt | |
6183 | << dendl; | |
6184 | } | |
6185 | } | |
6186 | if (cct != nullptr) { | |
6187 | ldout(cct,30) << __func__ << " pool " << pool_id | |
6188 | << " total_osd_weight " << total_osd_weight | |
6189 | << " total_weighted_pa " << total_weighted_pa | |
6190 | << dendl; | |
6191 | } | |
6192 | ||
6193 | if (prim_affinity_sum == 0.0) { | |
6194 | if (cct != nullptr) { | |
6195 | ldout(cct, 10) << __func__ << " pool " << pool_id | |
6196 | << " has primary_affinity set to zero on all OSDs" << dendl; | |
6197 | } | |
6198 | zero_rbi(*p_rbi); | |
6199 | p_rbi->err_msg = fmt::format("pool {} has primary_affinity set to zero on all OSDs", pool_id); | |
6200 | ||
6201 | return -ERANGE; // score has a different meaning now. | |
6202 | } | |
6203 | else { | |
6204 | max_osd_score *= prim_affinity_sum / num_osds; | |
6205 | } | |
6206 | ||
6207 | rc = tmp_osd_map.set_rbi(cct, *p_rbi, pool_id, total_weighted_pa, | |
6208 | prim_affinity_sum, num_osds, osd_pa_count, | |
6209 | total_osd_weight, max_prims_per_osd, | |
6210 | max_acting_prims_per_osd, avg_prims_per_osd, | |
6211 | prim_on_zero_pa, acting_on_zero_pa, max_osd_score); | |
6212 | ||
6213 | if (cct != nullptr) { | |
6214 | ldout(cct,30) << __func__ << " pool " << get_pool_name(pool_id) | |
6215 | << " pa_avg " << p_rbi->pa_avg | |
6216 | << " pa_weighted " << p_rbi->pa_weighted | |
6217 | << " pa_weighted_avg " << p_rbi->pa_weighted_avg | |
6218 | << " optimal_score " << p_rbi->optimal_score | |
6219 | << " adjusted_score " << p_rbi->adjusted_score | |
6220 | << " acting_adj_score " << p_rbi->acting_adj_score | |
6221 | << dendl; | |
6222 | ldout(cct,20) << __func__ << " pool " << get_pool_name(pool_id) | |
6223 | << " raw_score: " << p_rbi->raw_score | |
6224 | << " acting_raw_score: " << p_rbi->acting_raw_score | |
6225 | << dendl; | |
6226 | ldout(cct,10) << __func__ << " pool " << get_pool_name(pool_id) | |
6227 | << " wl_score: " << p_rbi->acting_adj_score << dendl; | |
6228 | } | |
6229 | ||
6230 | return rc; | |
6231 | } | |
6232 | ||
31f18b77 FG |
6233 | int OSDMap::get_osds_by_bucket_name(const string &name, set<int> *osds) const |
6234 | { | |
6235 | return crush->get_leaves(name, osds); | |
6236 | } | |
6237 | ||
3efd9988 FG |
6238 | // get pools whose crush rules might reference the given osd |
6239 | void OSDMap::get_pool_ids_by_osd(CephContext *cct, | |
6240 | int osd, | |
6241 | set<int64_t> *pool_ids) const | |
6242 | { | |
11fdf7f2 | 6243 | ceph_assert(pool_ids); |
3efd9988 FG |
6244 | set<int> raw_rules; |
6245 | int r = crush->get_rules_by_osd(osd, &raw_rules); | |
6246 | if (r < 0) { | |
6247 | lderr(cct) << __func__ << " get_rules_by_osd failed: " << cpp_strerror(r) | |
6248 | << dendl; | |
11fdf7f2 | 6249 | ceph_assert(r >= 0); |
3efd9988 FG |
6250 | } |
6251 | set<int> rules; | |
6252 | for (auto &i: raw_rules) { | |
6253 | // exclude any dead rule | |
6254 | if (crush_rule_in_use(i)) { | |
6255 | rules.insert(i); | |
6256 | } | |
6257 | } | |
6258 | for (auto &r: rules) { | |
6259 | get_pool_ids_by_rule(r, pool_ids); | |
6260 | } | |
6261 | } | |
6262 | ||
31f18b77 FG |
6263 | template <typename F> |
6264 | class OSDUtilizationDumper : public CrushTreeDumper::Dumper<F> { | |
6265 | public: | |
6266 | typedef CrushTreeDumper::Dumper<F> Parent; | |
6267 | ||
6268 | OSDUtilizationDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
11fdf7f2 | 6269 | const PGMap& pgmap_, bool tree_, |
9f95a23c | 6270 | const string& filter) : |
c07f9fc5 | 6271 | Parent(crush, osdmap_->get_pool_names()), |
31f18b77 | 6272 | osdmap(osdmap_), |
11fdf7f2 | 6273 | pgmap(pgmap_), |
31f18b77 | 6274 | tree(tree_), |
31f18b77 FG |
6275 | min_var(-1), |
6276 | max_var(-1), | |
6277 | stddev(0), | |
6278 | sum(0) { | |
9f95a23c TL |
6279 | if (osdmap->crush->name_exists(filter)) { |
6280 | // filter by crush node | |
6281 | auto item_id = osdmap->crush->get_item_id(filter); | |
11fdf7f2 TL |
6282 | allowed.insert(item_id); |
6283 | osdmap->crush->get_all_children(item_id, &allowed); | |
9f95a23c TL |
6284 | } else if (osdmap->crush->class_exists(filter)) { |
6285 | // filter by device class | |
6286 | class_id = osdmap->crush->get_class_id(filter); | |
6287 | } else if (auto pool_id = osdmap->lookup_pg_pool_name(filter); | |
6288 | pool_id >= 0) { | |
6289 | // filter by pool | |
6290 | auto crush_rule = osdmap->get_pool_crush_rule(pool_id); | |
6291 | set<int> roots; | |
6292 | osdmap->crush->find_takes_by_rule(crush_rule, &roots); | |
6293 | allowed = roots; | |
6294 | for (auto r : roots) | |
6295 | osdmap->crush->get_all_children(r, &allowed); | |
11fdf7f2 TL |
6296 | } |
6297 | average_util = average_utilization(); | |
31f18b77 FG |
6298 | } |
6299 | ||
6300 | protected: | |
11fdf7f2 TL |
6301 | |
6302 | bool should_dump(int id) const { | |
6303 | if (!allowed.empty() && !allowed.count(id)) // filter by name | |
6304 | return false; | |
9f95a23c TL |
6305 | if (id >= 0 && class_id >= 0) { |
6306 | auto item_class_id = osdmap->crush->get_item_class_id(id); | |
6307 | if (item_class_id < 0 || // not bound to a class yet | |
6308 | item_class_id != class_id) // or already bound to a different class | |
11fdf7f2 TL |
6309 | return false; |
6310 | } | |
6311 | return true; | |
6312 | } | |
6313 | ||
6314 | set<int> get_dumped_osds() { | |
9f95a23c | 6315 | if (allowed.empty() && class_id < 0) { |
11fdf7f2 TL |
6316 | // old way, all |
6317 | return {}; | |
6318 | } | |
6319 | return dumped_osds; | |
6320 | } | |
6321 | ||
31f18b77 FG |
6322 | void dump_stray(F *f) { |
6323 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
6324 | if (osdmap->exists(i) && !this->is_touched(i)) | |
c07f9fc5 | 6325 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
31f18b77 FG |
6326 | } |
6327 | } | |
6328 | ||
6329 | void dump_item(const CrushTreeDumper::Item &qi, F *f) override { | |
f67539c2 | 6330 | if (!tree && (qi.is_bucket() || dumped_osds.count(qi.id))) |
31f18b77 | 6331 | return; |
11fdf7f2 TL |
6332 | if (!should_dump(qi.id)) |
6333 | return; | |
31f18b77 | 6334 | |
11fdf7f2 TL |
6335 | if (!qi.is_bucket()) |
6336 | dumped_osds.insert(qi.id); | |
31f18b77 | 6337 | float reweight = qi.is_bucket() ? -1 : osdmap->get_weightf(qi.id); |
11fdf7f2 TL |
6338 | int64_t kb = 0, kb_used = 0, kb_used_data = 0, kb_used_omap = 0, |
6339 | kb_used_meta = 0, kb_avail = 0; | |
31f18b77 | 6340 | double util = 0; |
11fdf7f2 TL |
6341 | if (get_bucket_utilization(qi.id, &kb, &kb_used, &kb_used_data, |
6342 | &kb_used_omap, &kb_used_meta, &kb_avail)) | |
31f18b77 FG |
6343 | if (kb_used && kb) |
6344 | util = 100.0 * (double)kb_used / (double)kb; | |
6345 | ||
6346 | double var = 1.0; | |
6347 | if (average_util) | |
6348 | var = util / average_util; | |
6349 | ||
11fdf7f2 | 6350 | size_t num_pgs = qi.is_bucket() ? 0 : pgmap.get_num_pg_by_osd(qi.id); |
31f18b77 | 6351 | |
11fdf7f2 TL |
6352 | dump_item(qi, reweight, kb, kb_used, |
6353 | kb_used_data, kb_used_omap, kb_used_meta, | |
6354 | kb_avail, util, var, num_pgs, f); | |
31f18b77 FG |
6355 | |
6356 | if (!qi.is_bucket() && reweight > 0) { | |
6357 | if (min_var < 0 || var < min_var) | |
6358 | min_var = var; | |
6359 | if (max_var < 0 || var > max_var) | |
6360 | max_var = var; | |
6361 | ||
6362 | double dev = util - average_util; | |
6363 | dev *= dev; | |
6364 | stddev += reweight * dev; | |
6365 | sum += reweight; | |
6366 | } | |
6367 | } | |
6368 | ||
6369 | virtual void dump_item(const CrushTreeDumper::Item &qi, | |
6370 | float &reweight, | |
6371 | int64_t kb, | |
6372 | int64_t kb_used, | |
11fdf7f2 TL |
6373 | int64_t kb_used_data, |
6374 | int64_t kb_used_omap, | |
6375 | int64_t kb_used_meta, | |
31f18b77 FG |
6376 | int64_t kb_avail, |
6377 | double& util, | |
6378 | double& var, | |
6379 | const size_t num_pgs, | |
6380 | F *f) = 0; | |
6381 | ||
6382 | double dev() { | |
6383 | return sum > 0 ? sqrt(stddev / sum) : 0; | |
6384 | } | |
6385 | ||
6386 | double average_utilization() { | |
6387 | int64_t kb = 0, kb_used = 0; | |
6388 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
11fdf7f2 TL |
6389 | if (!osdmap->exists(i) || |
6390 | osdmap->get_weight(i) == 0 || | |
6391 | !should_dump(i)) | |
31f18b77 | 6392 | continue; |
11fdf7f2 TL |
6393 | int64_t kb_i, kb_used_i, kb_used_data_i, kb_used_omap_i, kb_used_meta_i, |
6394 | kb_avail_i; | |
6395 | if (get_osd_utilization(i, &kb_i, &kb_used_i, &kb_used_data_i, | |
6396 | &kb_used_omap_i, &kb_used_meta_i, &kb_avail_i)) { | |
31f18b77 FG |
6397 | kb += kb_i; |
6398 | kb_used += kb_used_i; | |
6399 | } | |
6400 | } | |
6401 | return kb > 0 ? 100.0 * (double)kb_used / (double)kb : 0; | |
6402 | } | |
6403 | ||
6404 | bool get_osd_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
6405 | int64_t* kb_used_data, |
6406 | int64_t* kb_used_omap, | |
6407 | int64_t* kb_used_meta, | |
31f18b77 | 6408 | int64_t* kb_avail) const { |
11fdf7f2 | 6409 | const osd_stat_t *p = pgmap.get_osd_stat(id); |
31f18b77 | 6410 | if (!p) return false; |
11fdf7f2 TL |
6411 | *kb = p->statfs.kb(); |
6412 | *kb_used = p->statfs.kb_used_raw(); | |
6413 | *kb_used_data = p->statfs.kb_used_data(); | |
6414 | *kb_used_omap = p->statfs.kb_used_omap(); | |
6415 | *kb_used_meta = p->statfs.kb_used_internal_metadata(); | |
6416 | *kb_avail = p->statfs.kb_avail(); | |
6417 | ||
f67539c2 | 6418 | return true; |
31f18b77 FG |
6419 | } |
6420 | ||
6421 | bool get_bucket_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
6422 | int64_t* kb_used_data, |
6423 | int64_t* kb_used_omap, | |
6424 | int64_t* kb_used_meta, | |
31f18b77 FG |
6425 | int64_t* kb_avail) const { |
6426 | if (id >= 0) { | |
11fdf7f2 | 6427 | if (osdmap->is_out(id) || !should_dump(id)) { |
31f18b77 FG |
6428 | *kb = 0; |
6429 | *kb_used = 0; | |
11fdf7f2 TL |
6430 | *kb_used_data = 0; |
6431 | *kb_used_omap = 0; | |
6432 | *kb_used_meta = 0; | |
31f18b77 FG |
6433 | *kb_avail = 0; |
6434 | return true; | |
6435 | } | |
11fdf7f2 TL |
6436 | return get_osd_utilization(id, kb, kb_used, kb_used_data, |
6437 | kb_used_omap, kb_used_meta, kb_avail); | |
31f18b77 FG |
6438 | } |
6439 | ||
6440 | *kb = 0; | |
6441 | *kb_used = 0; | |
11fdf7f2 TL |
6442 | *kb_used_data = 0; |
6443 | *kb_used_omap = 0; | |
6444 | *kb_used_meta = 0; | |
31f18b77 FG |
6445 | *kb_avail = 0; |
6446 | ||
6447 | for (int k = osdmap->crush->get_bucket_size(id) - 1; k >= 0; k--) { | |
6448 | int item = osdmap->crush->get_bucket_item(id, k); | |
11fdf7f2 TL |
6449 | int64_t kb_i = 0, kb_used_i = 0, kb_used_data_i = 0, |
6450 | kb_used_omap_i = 0, kb_used_meta_i = 0, kb_avail_i = 0; | |
6451 | if (!get_bucket_utilization(item, &kb_i, &kb_used_i, | |
6452 | &kb_used_data_i, &kb_used_omap_i, | |
6453 | &kb_used_meta_i, &kb_avail_i)) | |
31f18b77 FG |
6454 | return false; |
6455 | *kb += kb_i; | |
6456 | *kb_used += kb_used_i; | |
11fdf7f2 TL |
6457 | *kb_used_data += kb_used_data_i; |
6458 | *kb_used_omap += kb_used_omap_i; | |
6459 | *kb_used_meta += kb_used_meta_i; | |
31f18b77 FG |
6460 | *kb_avail += kb_avail_i; |
6461 | } | |
f67539c2 | 6462 | return true; |
31f18b77 FG |
6463 | } |
6464 | ||
6465 | protected: | |
6466 | const OSDMap *osdmap; | |
11fdf7f2 | 6467 | const PGMap& pgmap; |
31f18b77 FG |
6468 | bool tree; |
6469 | double average_util; | |
6470 | double min_var; | |
6471 | double max_var; | |
6472 | double stddev; | |
6473 | double sum; | |
9f95a23c | 6474 | int class_id = -1; |
11fdf7f2 TL |
6475 | set<int> allowed; |
6476 | set<int> dumped_osds; | |
31f18b77 FG |
6477 | }; |
6478 | ||
6479 | ||
6480 | class OSDUtilizationPlainDumper : public OSDUtilizationDumper<TextTable> { | |
6481 | public: | |
6482 | typedef OSDUtilizationDumper<TextTable> Parent; | |
6483 | ||
6484 | OSDUtilizationPlainDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 6485 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
6486 | const string& filter) : |
6487 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
6488 | |
6489 | void dump(TextTable *tbl) { | |
6490 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 6491 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
6492 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
6493 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); | |
6494 | tbl->define_column("SIZE", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 TL |
6495 | tbl->define_column("RAW USE", TextTable::LEFT, TextTable::RIGHT); |
6496 | tbl->define_column("DATA", TextTable::LEFT, TextTable::RIGHT); | |
6497 | tbl->define_column("OMAP", TextTable::LEFT, TextTable::RIGHT); | |
6498 | tbl->define_column("META", TextTable::LEFT, TextTable::RIGHT); | |
31f18b77 FG |
6499 | tbl->define_column("AVAIL", TextTable::LEFT, TextTable::RIGHT); |
6500 | tbl->define_column("%USE", TextTable::LEFT, TextTable::RIGHT); | |
6501 | tbl->define_column("VAR", TextTable::LEFT, TextTable::RIGHT); | |
6502 | tbl->define_column("PGS", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 | 6503 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
6504 | if (tree) |
6505 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
6506 | ||
6507 | Parent::dump(tbl); | |
6508 | ||
6509 | dump_stray(tbl); | |
6510 | ||
11fdf7f2 | 6511 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
224ce89b WB |
6512 | *tbl << "" |
6513 | << "" | |
6514 | << "" << "TOTAL" | |
11fdf7f2 TL |
6515 | << byte_u_t(sum.statfs.total) |
6516 | << byte_u_t(sum.statfs.get_used_raw()) | |
6517 | << byte_u_t(sum.statfs.allocated) | |
6518 | << byte_u_t(sum.statfs.omap_allocated) | |
6519 | << byte_u_t(sum.statfs.internal_metadata) | |
6520 | << byte_u_t(sum.statfs.available) | |
31f18b77 FG |
6521 | << lowprecision_t(average_util) |
6522 | << "" | |
6523 | << TextTable::endrow; | |
6524 | } | |
6525 | ||
6526 | protected: | |
6527 | struct lowprecision_t { | |
6528 | float v; | |
6529 | explicit lowprecision_t(float _v) : v(_v) {} | |
6530 | }; | |
6531 | friend std::ostream &operator<<(ostream& out, const lowprecision_t& v); | |
6532 | ||
6533 | using OSDUtilizationDumper<TextTable>::dump_item; | |
6534 | void dump_item(const CrushTreeDumper::Item &qi, | |
6535 | float &reweight, | |
6536 | int64_t kb, | |
6537 | int64_t kb_used, | |
11fdf7f2 TL |
6538 | int64_t kb_used_data, |
6539 | int64_t kb_used_omap, | |
6540 | int64_t kb_used_meta, | |
31f18b77 FG |
6541 | int64_t kb_avail, |
6542 | double& util, | |
6543 | double& var, | |
6544 | const size_t num_pgs, | |
6545 | TextTable *tbl) override { | |
224ce89b WB |
6546 | const char *c = crush->get_item_class(qi.id); |
6547 | if (!c) | |
6548 | c = ""; | |
31f18b77 | 6549 | *tbl << qi.id |
224ce89b | 6550 | << c |
31f18b77 FG |
6551 | << weightf_t(qi.weight) |
6552 | << weightf_t(reweight) | |
1adf2230 AA |
6553 | << byte_u_t(kb << 10) |
6554 | << byte_u_t(kb_used << 10) | |
11fdf7f2 TL |
6555 | << byte_u_t(kb_used_data << 10) |
6556 | << byte_u_t(kb_used_omap << 10) | |
6557 | << byte_u_t(kb_used_meta << 10) | |
1adf2230 | 6558 | << byte_u_t(kb_avail << 10) |
31f18b77 FG |
6559 | << lowprecision_t(util) |
6560 | << lowprecision_t(var); | |
6561 | ||
6562 | if (qi.is_bucket()) { | |
6563 | *tbl << "-"; | |
11fdf7f2 | 6564 | *tbl << ""; |
31f18b77 FG |
6565 | } else { |
6566 | *tbl << num_pgs; | |
11fdf7f2 TL |
6567 | if (osdmap->is_up(qi.id)) { |
6568 | *tbl << "up"; | |
6569 | } else if (osdmap->is_destroyed(qi.id)) { | |
6570 | *tbl << "destroyed"; | |
6571 | } else { | |
6572 | *tbl << "down"; | |
6573 | } | |
31f18b77 FG |
6574 | } |
6575 | ||
6576 | if (tree) { | |
6577 | ostringstream name; | |
6578 | for (int k = 0; k < qi.depth; k++) | |
6579 | name << " "; | |
6580 | if (qi.is_bucket()) { | |
6581 | int type = crush->get_bucket_type(qi.id); | |
6582 | name << crush->get_type_name(type) << " " | |
6583 | << crush->get_item_name(qi.id); | |
6584 | } else { | |
6585 | name << "osd." << qi.id; | |
6586 | } | |
6587 | *tbl << name.str(); | |
6588 | } | |
6589 | ||
6590 | *tbl << TextTable::endrow; | |
6591 | } | |
6592 | ||
6593 | public: | |
6594 | string summary() { | |
6595 | ostringstream out; | |
6596 | out << "MIN/MAX VAR: " << lowprecision_t(min_var) | |
6597 | << "/" << lowprecision_t(max_var) << " " | |
6598 | << "STDDEV: " << lowprecision_t(dev()); | |
6599 | return out.str(); | |
6600 | } | |
6601 | }; | |
6602 | ||
6603 | ostream& operator<<(ostream& out, | |
6604 | const OSDUtilizationPlainDumper::lowprecision_t& v) | |
6605 | { | |
6606 | if (v.v < -0.01) { | |
6607 | return out << "-"; | |
6608 | } else if (v.v < 0.001) { | |
6609 | return out << "0"; | |
6610 | } else { | |
6611 | std::streamsize p = out.precision(); | |
6612 | return out << std::fixed << std::setprecision(2) << v.v << std::setprecision(p); | |
6613 | } | |
6614 | } | |
6615 | ||
6616 | class OSDUtilizationFormatDumper : public OSDUtilizationDumper<Formatter> { | |
6617 | public: | |
6618 | typedef OSDUtilizationDumper<Formatter> Parent; | |
6619 | ||
6620 | OSDUtilizationFormatDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 6621 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
6622 | const string& filter) : |
6623 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
6624 | |
6625 | void dump(Formatter *f) { | |
6626 | f->open_array_section("nodes"); | |
6627 | Parent::dump(f); | |
6628 | f->close_section(); | |
6629 | ||
6630 | f->open_array_section("stray"); | |
6631 | dump_stray(f); | |
6632 | f->close_section(); | |
6633 | } | |
6634 | ||
6635 | protected: | |
6636 | using OSDUtilizationDumper<Formatter>::dump_item; | |
6637 | void dump_item(const CrushTreeDumper::Item &qi, | |
11fdf7f2 TL |
6638 | float &reweight, |
6639 | int64_t kb, | |
6640 | int64_t kb_used, | |
6641 | int64_t kb_used_data, | |
6642 | int64_t kb_used_omap, | |
6643 | int64_t kb_used_meta, | |
6644 | int64_t kb_avail, | |
6645 | double& util, | |
6646 | double& var, | |
6647 | const size_t num_pgs, | |
6648 | Formatter *f) override { | |
31f18b77 | 6649 | f->open_object_section("item"); |
c07f9fc5 | 6650 | CrushTreeDumper::dump_item_fields(crush, weight_set_names, qi, f); |
31f18b77 FG |
6651 | f->dump_float("reweight", reweight); |
6652 | f->dump_int("kb", kb); | |
6653 | f->dump_int("kb_used", kb_used); | |
11fdf7f2 TL |
6654 | f->dump_int("kb_used_data", kb_used_data); |
6655 | f->dump_int("kb_used_omap", kb_used_omap); | |
6656 | f->dump_int("kb_used_meta", kb_used_meta); | |
31f18b77 FG |
6657 | f->dump_int("kb_avail", kb_avail); |
6658 | f->dump_float("utilization", util); | |
6659 | f->dump_float("var", var); | |
6660 | f->dump_unsigned("pgs", num_pgs); | |
11fdf7f2 TL |
6661 | if (!qi.is_bucket()) { |
6662 | if (osdmap->is_up(qi.id)) { | |
6663 | f->dump_string("status", "up"); | |
6664 | } else if (osdmap->is_destroyed(qi.id)) { | |
6665 | f->dump_string("status", "destroyed"); | |
6666 | } else { | |
6667 | f->dump_string("status", "down"); | |
6668 | } | |
6669 | } | |
31f18b77 FG |
6670 | CrushTreeDumper::dump_bucket_children(crush, qi, f); |
6671 | f->close_section(); | |
6672 | } | |
6673 | ||
6674 | public: | |
6675 | void summary(Formatter *f) { | |
6676 | f->open_object_section("summary"); | |
11fdf7f2 TL |
6677 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
6678 | auto& s = sum.statfs; | |
6679 | ||
6680 | f->dump_int("total_kb", s.kb()); | |
6681 | f->dump_int("total_kb_used", s.kb_used_raw()); | |
6682 | f->dump_int("total_kb_used_data", s.kb_used_data()); | |
6683 | f->dump_int("total_kb_used_omap", s.kb_used_omap()); | |
6684 | f->dump_int("total_kb_used_meta", s.kb_used_internal_metadata()); | |
6685 | f->dump_int("total_kb_avail", s.kb_avail()); | |
31f18b77 FG |
6686 | f->dump_float("average_utilization", average_util); |
6687 | f->dump_float("min_var", min_var); | |
6688 | f->dump_float("max_var", max_var); | |
6689 | f->dump_float("dev", dev()); | |
6690 | f->close_section(); | |
6691 | } | |
6692 | }; | |
6693 | ||
6694 | void print_osd_utilization(const OSDMap& osdmap, | |
11fdf7f2 TL |
6695 | const PGMap& pgmap, |
6696 | ostream& out, | |
6697 | Formatter *f, | |
6698 | bool tree, | |
9f95a23c | 6699 | const string& filter) |
31f18b77 FG |
6700 | { |
6701 | const CrushWrapper *crush = osdmap.crush.get(); | |
6702 | if (f) { | |
6703 | f->open_object_section("df"); | |
9f95a23c | 6704 | OSDUtilizationFormatDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
6705 | d.dump(f); |
6706 | d.summary(f); | |
6707 | f->close_section(); | |
6708 | f->flush(out); | |
6709 | } else { | |
9f95a23c | 6710 | OSDUtilizationPlainDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
6711 | TextTable tbl; |
6712 | d.dump(&tbl); | |
6713 | out << tbl << d.summary() << "\n"; | |
6714 | } | |
6715 | } | |
224ce89b | 6716 | |
92f5a8d4 TL |
6717 | void OSDMap::check_health(CephContext *cct, |
6718 | health_check_map_t *checks) const | |
224ce89b WB |
6719 | { |
6720 | int num_osds = get_num_osds(); | |
6721 | ||
6722 | // OSD_DOWN | |
6723 | // OSD_$subtree_DOWN | |
6724 | // OSD_ORPHAN | |
6725 | if (num_osds >= 0) { | |
6726 | int num_in_osds = 0; | |
6727 | int num_down_in_osds = 0; | |
6728 | set<int> osds; | |
6729 | set<int> down_in_osds; | |
6730 | set<int> up_in_osds; | |
6731 | set<int> subtree_up; | |
6732 | unordered_map<int, set<int> > subtree_type_down; | |
6733 | unordered_map<int, int> num_osds_subtree; | |
6734 | int max_type = crush->get_max_type_id(); | |
6735 | ||
6736 | for (int i = 0; i < get_max_osd(); i++) { | |
6737 | if (!exists(i)) { | |
6738 | if (crush->item_exists(i)) { | |
6739 | osds.insert(i); | |
6740 | } | |
6741 | continue; | |
6742 | } | |
f67539c2 | 6743 | if (is_out(i) || (osd_state[i] & CEPH_OSD_NEW)) |
224ce89b WB |
6744 | continue; |
6745 | ++num_in_osds; | |
6746 | if (down_in_osds.count(i) || up_in_osds.count(i)) | |
6747 | continue; | |
6748 | if (!is_up(i)) { | |
6749 | down_in_osds.insert(i); | |
6750 | int parent_id = 0; | |
6751 | int current = i; | |
6752 | for (int type = 0; type <= max_type; type++) { | |
6753 | if (!crush->get_type_name(type)) | |
6754 | continue; | |
6755 | int r = crush->get_immediate_parent_id(current, &parent_id); | |
6756 | if (r == -ENOENT) | |
6757 | break; | |
6758 | // break early if this parent is already marked as up | |
6759 | if (subtree_up.count(parent_id)) | |
6760 | break; | |
6761 | type = crush->get_bucket_type(parent_id); | |
6762 | if (!subtree_type_is_down( | |
92f5a8d4 | 6763 | cct, parent_id, type, |
224ce89b WB |
6764 | &down_in_osds, &up_in_osds, &subtree_up, &subtree_type_down)) |
6765 | break; | |
6766 | current = parent_id; | |
6767 | } | |
6768 | } | |
6769 | } | |
6770 | ||
6771 | // calculate the number of down osds in each down subtree and | |
6772 | // store it in num_osds_subtree | |
6773 | for (int type = 1; type <= max_type; type++) { | |
6774 | if (!crush->get_type_name(type)) | |
6775 | continue; | |
6776 | for (auto j = subtree_type_down[type].begin(); | |
6777 | j != subtree_type_down[type].end(); | |
6778 | ++j) { | |
6779 | list<int> children; | |
6780 | int num = 0; | |
6781 | int num_children = crush->get_children(*j, &children); | |
6782 | if (num_children == 0) | |
6783 | continue; | |
6784 | for (auto l = children.begin(); l != children.end(); ++l) { | |
6785 | if (*l >= 0) { | |
6786 | ++num; | |
6787 | } else if (num_osds_subtree[*l] > 0) { | |
6788 | num = num + num_osds_subtree[*l]; | |
6789 | } | |
6790 | } | |
6791 | num_osds_subtree[*j] = num; | |
6792 | } | |
6793 | } | |
6794 | num_down_in_osds = down_in_osds.size(); | |
11fdf7f2 | 6795 | ceph_assert(num_down_in_osds <= num_in_osds); |
224ce89b WB |
6796 | if (num_down_in_osds > 0) { |
6797 | // summary of down subtree types and osds | |
6798 | for (int type = max_type; type > 0; type--) { | |
6799 | if (!crush->get_type_name(type)) | |
6800 | continue; | |
6801 | if (subtree_type_down[type].size() > 0) { | |
6802 | ostringstream ss; | |
6803 | ss << subtree_type_down[type].size() << " " | |
6804 | << crush->get_type_name(type); | |
6805 | if (subtree_type_down[type].size() > 1) { | |
6806 | ss << "s"; | |
6807 | } | |
6808 | int sum_down_osds = 0; | |
6809 | for (auto j = subtree_type_down[type].begin(); | |
6810 | j != subtree_type_down[type].end(); | |
6811 | ++j) { | |
6812 | sum_down_osds = sum_down_osds + num_osds_subtree[*j]; | |
6813 | } | |
6814 | ss << " (" << sum_down_osds << " osds) down"; | |
6815 | string err = string("OSD_") + | |
6816 | string(crush->get_type_name(type)) + "_DOWN"; | |
6817 | boost::to_upper(err); | |
9f95a23c TL |
6818 | auto& d = checks->add(err, HEALTH_WARN, ss.str(), |
6819 | subtree_type_down[type].size()); | |
224ce89b WB |
6820 | for (auto j = subtree_type_down[type].rbegin(); |
6821 | j != subtree_type_down[type].rend(); | |
6822 | ++j) { | |
6823 | ostringstream ss; | |
6824 | ss << crush->get_type_name(type); | |
6825 | ss << " "; | |
6826 | ss << crush->get_item_name(*j); | |
6827 | // at the top level, do not print location | |
6828 | if (type != max_type) { | |
6829 | ss << " ("; | |
6830 | ss << crush->get_full_location_ordered_string(*j); | |
6831 | ss << ")"; | |
6832 | } | |
6833 | int num = num_osds_subtree[*j]; | |
6834 | ss << " (" << num << " osds)"; | |
6835 | ss << " is down"; | |
6836 | d.detail.push_back(ss.str()); | |
6837 | } | |
6838 | } | |
6839 | } | |
6840 | ostringstream ss; | |
6841 | ss << down_in_osds.size() << " osds down"; | |
9f95a23c TL |
6842 | auto& d = checks->add("OSD_DOWN", HEALTH_WARN, ss.str(), |
6843 | down_in_osds.size()); | |
224ce89b WB |
6844 | for (auto it = down_in_osds.begin(); it != down_in_osds.end(); ++it) { |
6845 | ostringstream ss; | |
6846 | ss << "osd." << *it << " ("; | |
6847 | ss << crush->get_full_location_ordered_string(*it); | |
6848 | ss << ") is down"; | |
6849 | d.detail.push_back(ss.str()); | |
6850 | } | |
6851 | } | |
6852 | ||
6853 | if (!osds.empty()) { | |
6854 | ostringstream ss; | |
6855 | ss << osds.size() << " osds exist in the crush map but not in the osdmap"; | |
9f95a23c TL |
6856 | auto& d = checks->add("OSD_ORPHAN", HEALTH_WARN, ss.str(), |
6857 | osds.size()); | |
224ce89b WB |
6858 | for (auto osd : osds) { |
6859 | ostringstream ss; | |
6860 | ss << "osd." << osd << " exists in crush map but not in osdmap"; | |
6861 | d.detail.push_back(ss.str()); | |
6862 | } | |
6863 | } | |
6864 | } | |
6865 | ||
eafe8130 TL |
6866 | std::list<std::string> scrub_messages; |
6867 | bool noscrub = false, nodeepscrub = false; | |
6868 | for (const auto &p : pools) { | |
6869 | if (p.second.flags & pg_pool_t::FLAG_NOSCRUB) { | |
6870 | ostringstream ss; | |
6871 | ss << "Pool " << get_pool_name(p.first) << " has noscrub flag"; | |
6872 | scrub_messages.push_back(ss.str()); | |
6873 | noscrub = true; | |
6874 | } | |
6875 | if (p.second.flags & pg_pool_t::FLAG_NODEEP_SCRUB) { | |
6876 | ostringstream ss; | |
6877 | ss << "Pool " << get_pool_name(p.first) << " has nodeep-scrub flag"; | |
6878 | scrub_messages.push_back(ss.str()); | |
6879 | nodeepscrub = true; | |
6880 | } | |
6881 | } | |
6882 | if (noscrub || nodeepscrub) { | |
6883 | string out = ""; | |
6884 | out += noscrub ? string("noscrub") + (nodeepscrub ? ", " : "") : ""; | |
6885 | out += nodeepscrub ? "nodeep-scrub" : ""; | |
6886 | auto& d = checks->add("POOL_SCRUB_FLAGS", HEALTH_OK, | |
9f95a23c | 6887 | "Some pool(s) have the " + out + " flag(s) set", 0); |
eafe8130 TL |
6888 | d.detail.splice(d.detail.end(), scrub_messages); |
6889 | } | |
6890 | ||
224ce89b WB |
6891 | // OSD_OUT_OF_ORDER_FULL |
6892 | { | |
6893 | // An osd could configure failsafe ratio, to something different | |
6894 | // but for now assume it is the same here. | |
92f5a8d4 | 6895 | float fsr = cct->_conf->osd_failsafe_full_ratio; |
224ce89b WB |
6896 | if (fsr > 1.0) fsr /= 100; |
6897 | float fr = get_full_ratio(); | |
6898 | float br = get_backfillfull_ratio(); | |
6899 | float nr = get_nearfull_ratio(); | |
6900 | ||
6901 | list<string> detail; | |
6902 | // These checks correspond to how OSDService::check_full_status() in an OSD | |
6903 | // handles the improper setting of these values. | |
6904 | if (br < nr) { | |
6905 | ostringstream ss; | |
6906 | ss << "backfillfull_ratio (" << br | |
6907 | << ") < nearfull_ratio (" << nr << "), increased"; | |
6908 | detail.push_back(ss.str()); | |
6909 | br = nr; | |
6910 | } | |
6911 | if (fr < br) { | |
6912 | ostringstream ss; | |
6913 | ss << "full_ratio (" << fr << ") < backfillfull_ratio (" << br | |
6914 | << "), increased"; | |
6915 | detail.push_back(ss.str()); | |
6916 | fr = br; | |
6917 | } | |
6918 | if (fsr < fr) { | |
6919 | ostringstream ss; | |
6920 | ss << "osd_failsafe_full_ratio (" << fsr << ") < full_ratio (" << fr | |
6921 | << "), increased"; | |
6922 | detail.push_back(ss.str()); | |
6923 | } | |
6924 | if (!detail.empty()) { | |
6925 | auto& d = checks->add("OSD_OUT_OF_ORDER_FULL", HEALTH_ERR, | |
9f95a23c | 6926 | "full ratio(s) out of order", 0); |
224ce89b WB |
6927 | d.detail.swap(detail); |
6928 | } | |
6929 | } | |
6930 | ||
6931 | // OSD_FULL | |
6932 | // OSD_NEARFULL | |
6933 | // OSD_BACKFILLFULL | |
6934 | // OSD_FAILSAFE_FULL | |
6935 | { | |
6936 | set<int> full, backfillfull, nearfull; | |
6937 | get_full_osd_counts(&full, &backfillfull, &nearfull); | |
6938 | if (full.size()) { | |
6939 | ostringstream ss; | |
6940 | ss << full.size() << " full osd(s)"; | |
9f95a23c | 6941 | auto& d = checks->add("OSD_FULL", HEALTH_ERR, ss.str(), full.size()); |
224ce89b WB |
6942 | for (auto& i: full) { |
6943 | ostringstream ss; | |
6944 | ss << "osd." << i << " is full"; | |
6945 | d.detail.push_back(ss.str()); | |
6946 | } | |
6947 | } | |
6948 | if (backfillfull.size()) { | |
6949 | ostringstream ss; | |
6950 | ss << backfillfull.size() << " backfillfull osd(s)"; | |
9f95a23c TL |
6951 | auto& d = checks->add("OSD_BACKFILLFULL", HEALTH_WARN, ss.str(), |
6952 | backfillfull.size()); | |
224ce89b WB |
6953 | for (auto& i: backfillfull) { |
6954 | ostringstream ss; | |
6955 | ss << "osd." << i << " is backfill full"; | |
6956 | d.detail.push_back(ss.str()); | |
6957 | } | |
6958 | } | |
6959 | if (nearfull.size()) { | |
6960 | ostringstream ss; | |
6961 | ss << nearfull.size() << " nearfull osd(s)"; | |
9f95a23c | 6962 | auto& d = checks->add("OSD_NEARFULL", HEALTH_WARN, ss.str(), nearfull.size()); |
224ce89b WB |
6963 | for (auto& i: nearfull) { |
6964 | ostringstream ss; | |
6965 | ss << "osd." << i << " is near full"; | |
6966 | d.detail.push_back(ss.str()); | |
6967 | } | |
6968 | } | |
6969 | } | |
6970 | ||
6971 | // OSDMAP_FLAGS | |
6972 | { | |
6973 | // warn about flags | |
6974 | uint64_t warn_flags = | |
224ce89b WB |
6975 | CEPH_OSDMAP_PAUSERD | |
6976 | CEPH_OSDMAP_PAUSEWR | | |
6977 | CEPH_OSDMAP_PAUSEREC | | |
6978 | CEPH_OSDMAP_NOUP | | |
6979 | CEPH_OSDMAP_NODOWN | | |
6980 | CEPH_OSDMAP_NOIN | | |
6981 | CEPH_OSDMAP_NOOUT | | |
6982 | CEPH_OSDMAP_NOBACKFILL | | |
6983 | CEPH_OSDMAP_NORECOVER | | |
6984 | CEPH_OSDMAP_NOSCRUB | | |
6985 | CEPH_OSDMAP_NODEEP_SCRUB | | |
6986 | CEPH_OSDMAP_NOTIERAGENT | | |
11fdf7f2 | 6987 | CEPH_OSDMAP_NOSNAPTRIM | |
224ce89b WB |
6988 | CEPH_OSDMAP_NOREBALANCE; |
6989 | if (test_flag(warn_flags)) { | |
6990 | ostringstream ss; | |
9f95a23c TL |
6991 | string s = get_flag_string(get_flags() & warn_flags); |
6992 | ss << s << " flag(s) set"; | |
6993 | checks->add("OSDMAP_FLAGS", HEALTH_WARN, ss.str(), | |
6994 | s.size() /* kludgey but sufficient */); | |
224ce89b WB |
6995 | } |
6996 | } | |
6997 | ||
6998 | // OSD_FLAGS | |
6999 | { | |
7000 | list<string> detail; | |
7001 | const unsigned flags = | |
7002 | CEPH_OSD_NOUP | | |
7003 | CEPH_OSD_NOIN | | |
7004 | CEPH_OSD_NODOWN | | |
7005 | CEPH_OSD_NOOUT; | |
7006 | for (int i = 0; i < max_osd; ++i) { | |
7007 | if (osd_state[i] & flags) { | |
7008 | ostringstream ss; | |
7009 | set<string> states; | |
7010 | OSDMap::calc_state_set(osd_state[i] & flags, states); | |
7011 | ss << "osd." << i << " has flags " << states; | |
7012 | detail.push_back(ss.str()); | |
7013 | } | |
7014 | } | |
81eedcae TL |
7015 | for (auto& i : crush_node_flags) { |
7016 | if (i.second && crush->item_exists(i.first)) { | |
7017 | ostringstream ss; | |
7018 | set<string> states; | |
7019 | OSDMap::calc_state_set(i.second, states); | |
7020 | int t = i.first >= 0 ? 0 : crush->get_bucket_type(i.first); | |
7021 | const char *tn = crush->get_type_name(t); | |
7022 | ss << (tn ? tn : "node") << " " | |
7023 | << crush->get_item_name(i.first) << " has flags " << states; | |
7024 | detail.push_back(ss.str()); | |
7025 | } | |
7026 | } | |
7027 | for (auto& i : device_class_flags) { | |
7028 | const char* class_name = crush->get_class_name(i.first); | |
7029 | if (i.second && class_name) { | |
7030 | ostringstream ss; | |
7031 | set<string> states; | |
7032 | OSDMap::calc_state_set(i.second, states); | |
7033 | ss << "device class '" << class_name << "' has flags " << states; | |
7034 | detail.push_back(ss.str()); | |
7035 | } | |
7036 | } | |
224ce89b WB |
7037 | if (!detail.empty()) { |
7038 | ostringstream ss; | |
81eedcae | 7039 | ss << detail.size() << " OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set"; |
9f95a23c | 7040 | auto& d = checks->add("OSD_FLAGS", HEALTH_WARN, ss.str(), detail.size()); |
224ce89b WB |
7041 | d.detail.swap(detail); |
7042 | } | |
7043 | } | |
7044 | ||
7045 | // OLD_CRUSH_TUNABLES | |
92f5a8d4 | 7046 | if (cct->_conf->mon_warn_on_legacy_crush_tunables) { |
224ce89b | 7047 | string min = crush->get_min_required_version(); |
92f5a8d4 | 7048 | if (min < cct->_conf->mon_crush_min_required_version) { |
224ce89b WB |
7049 | ostringstream ss; |
7050 | ss << "crush map has legacy tunables (require " << min | |
92f5a8d4 | 7051 | << ", min is " << cct->_conf->mon_crush_min_required_version << ")"; |
9f95a23c | 7052 | auto& d = checks->add("OLD_CRUSH_TUNABLES", HEALTH_WARN, ss.str(), 0); |
f67539c2 | 7053 | d.detail.push_back("see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
7054 | } |
7055 | } | |
7056 | ||
7057 | // OLD_CRUSH_STRAW_CALC_VERSION | |
92f5a8d4 | 7058 | if (cct->_conf->mon_warn_on_crush_straw_calc_version_zero) { |
224ce89b WB |
7059 | if (crush->get_straw_calc_version() == 0) { |
7060 | ostringstream ss; | |
7061 | ss << "crush map has straw_calc_version=0"; | |
9f95a23c | 7062 | auto& d = checks->add("OLD_CRUSH_STRAW_CALC_VERSION", HEALTH_WARN, ss.str(), 0); |
224ce89b | 7063 | d.detail.push_back( |
f67539c2 | 7064 | "see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
7065 | } |
7066 | } | |
7067 | ||
7068 | // CACHE_POOL_NO_HIT_SET | |
92f5a8d4 | 7069 | if (cct->_conf->mon_warn_on_cache_pools_without_hit_sets) { |
224ce89b | 7070 | list<string> detail; |
9f95a23c | 7071 | for (auto p = pools.cbegin(); p != pools.cend(); ++p) { |
224ce89b WB |
7072 | const pg_pool_t& info = p->second; |
7073 | if (info.cache_mode_requires_hit_set() && | |
7074 | info.hit_set_params.get_type() == HitSet::TYPE_NONE) { | |
7075 | ostringstream ss; | |
7076 | ss << "pool '" << get_pool_name(p->first) | |
7077 | << "' with cache_mode " << info.get_cache_mode_name() | |
7078 | << " needs hit_set_type to be set but it is not"; | |
7079 | detail.push_back(ss.str()); | |
7080 | } | |
7081 | } | |
7082 | if (!detail.empty()) { | |
7083 | ostringstream ss; | |
7084 | ss << detail.size() << " cache pools are missing hit_sets"; | |
9f95a23c TL |
7085 | auto& d = checks->add("CACHE_POOL_NO_HIT_SET", HEALTH_WARN, ss.str(), |
7086 | detail.size()); | |
224ce89b WB |
7087 | d.detail.swap(detail); |
7088 | } | |
7089 | } | |
7090 | ||
7091 | // OSD_NO_SORTBITWISE | |
11fdf7f2 | 7092 | if (!test_flag(CEPH_OSDMAP_SORTBITWISE)) { |
224ce89b | 7093 | ostringstream ss; |
11fdf7f2 | 7094 | ss << "'sortbitwise' flag is not set"; |
9f95a23c | 7095 | checks->add("OSD_NO_SORTBITWISE", HEALTH_WARN, ss.str(), 0); |
224ce89b WB |
7096 | } |
7097 | ||
7098 | // OSD_UPGRADE_FINISHED | |
20effc67 TL |
7099 | if (auto require_release = pending_require_osd_release()) { |
7100 | ostringstream ss; | |
7101 | ss << "all OSDs are running " << *require_release << " or later but" | |
7102 | << " require_osd_release < " << *require_release; | |
7103 | auto& d = checks->add("OSD_UPGRADE_FINISHED", HEALTH_WARN, ss.str(), 0); | |
7104 | d.detail.push_back(ss.str()); | |
7105 | } | |
224ce89b | 7106 | |
3efd9988 | 7107 | // POOL_NEARFULL/BACKFILLFULL/FULL |
224ce89b | 7108 | { |
3efd9988 | 7109 | list<string> full_detail, backfillfull_detail, nearfull_detail; |
224ce89b WB |
7110 | for (auto it : get_pools()) { |
7111 | const pg_pool_t &pool = it.second; | |
3efd9988 | 7112 | const string& pool_name = get_pool_name(it.first); |
224ce89b | 7113 | if (pool.has_flag(pg_pool_t::FLAG_FULL)) { |
224ce89b | 7114 | stringstream ss; |
11fdf7f2 | 7115 | if (pool.has_flag(pg_pool_t::FLAG_FULL_QUOTA)) { |
3efd9988 FG |
7116 | // may run out of space too, |
7117 | // but we want EQUOTA taking precedence | |
11fdf7f2 | 7118 | ss << "pool '" << pool_name << "' is full (running out of quota)"; |
3efd9988 FG |
7119 | } else { |
7120 | ss << "pool '" << pool_name << "' is full (no space)"; | |
7121 | } | |
7122 | full_detail.push_back(ss.str()); | |
7123 | } else if (pool.has_flag(pg_pool_t::FLAG_BACKFILLFULL)) { | |
7124 | stringstream ss; | |
7125 | ss << "pool '" << pool_name << "' is backfillfull"; | |
7126 | backfillfull_detail.push_back(ss.str()); | |
7127 | } else if (pool.has_flag(pg_pool_t::FLAG_NEARFULL)) { | |
7128 | stringstream ss; | |
7129 | ss << "pool '" << pool_name << "' is nearfull"; | |
7130 | nearfull_detail.push_back(ss.str()); | |
224ce89b WB |
7131 | } |
7132 | } | |
3efd9988 | 7133 | if (!full_detail.empty()) { |
224ce89b | 7134 | ostringstream ss; |
3efd9988 | 7135 | ss << full_detail.size() << " pool(s) full"; |
9f95a23c | 7136 | auto& d = checks->add("POOL_FULL", HEALTH_WARN, ss.str(), full_detail.size()); |
3efd9988 FG |
7137 | d.detail.swap(full_detail); |
7138 | } | |
7139 | if (!backfillfull_detail.empty()) { | |
7140 | ostringstream ss; | |
7141 | ss << backfillfull_detail.size() << " pool(s) backfillfull"; | |
9f95a23c TL |
7142 | auto& d = checks->add("POOL_BACKFILLFULL", HEALTH_WARN, ss.str(), |
7143 | backfillfull_detail.size()); | |
3efd9988 FG |
7144 | d.detail.swap(backfillfull_detail); |
7145 | } | |
7146 | if (!nearfull_detail.empty()) { | |
7147 | ostringstream ss; | |
7148 | ss << nearfull_detail.size() << " pool(s) nearfull"; | |
9f95a23c TL |
7149 | auto& d = checks->add("POOL_NEARFULL", HEALTH_WARN, ss.str(), |
7150 | nearfull_detail.size()); | |
3efd9988 | 7151 | d.detail.swap(nearfull_detail); |
224ce89b WB |
7152 | } |
7153 | } | |
92f5a8d4 TL |
7154 | |
7155 | // POOL_PG_NUM_NOT_POWER_OF_TWO | |
7156 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_pg_num_not_power_of_two")) { | |
7157 | list<string> detail; | |
7158 | for (auto it : get_pools()) { | |
1e59de90 | 7159 | if (!std::has_single_bit(it.second.get_pg_num_target())) { |
92f5a8d4 TL |
7160 | ostringstream ss; |
7161 | ss << "pool '" << get_pool_name(it.first) | |
7162 | << "' pg_num " << it.second.get_pg_num_target() | |
7163 | << " is not a power of two"; | |
7164 | detail.push_back(ss.str()); | |
7165 | } | |
7166 | } | |
7167 | if (!detail.empty()) { | |
7168 | ostringstream ss; | |
7169 | ss << detail.size() << " pool(s) have non-power-of-two pg_num"; | |
7170 | auto& d = checks->add("POOL_PG_NUM_NOT_POWER_OF_TWO", HEALTH_WARN, | |
9f95a23c TL |
7171 | ss.str(), detail.size()); |
7172 | d.detail.swap(detail); | |
7173 | } | |
7174 | } | |
7175 | ||
7176 | // POOL_NO_REDUNDANCY | |
7177 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_no_redundancy")) | |
7178 | { | |
7179 | list<string> detail; | |
7180 | for (auto it : get_pools()) { | |
7181 | if (it.second.get_size() == 1) { | |
7182 | ostringstream ss; | |
7183 | ss << "pool '" << get_pool_name(it.first) | |
7184 | << "' has no replicas configured"; | |
7185 | detail.push_back(ss.str()); | |
7186 | } | |
7187 | } | |
7188 | if (!detail.empty()) { | |
7189 | ostringstream ss; | |
7190 | ss << detail.size() << " pool(s) have no replicas configured"; | |
7191 | auto& d = checks->add("POOL_NO_REDUNDANCY", HEALTH_WARN, | |
7192 | ss.str(), detail.size()); | |
92f5a8d4 TL |
7193 | d.detail.swap(detail); |
7194 | } | |
7195 | } | |
f67539c2 TL |
7196 | |
7197 | // DEGRADED STRETCH MODE | |
7198 | if (cct->_conf.get_val<bool>("mon_warn_on_degraded_stretch_mode")) { | |
7199 | if (recovering_stretch_mode) { | |
7200 | stringstream ss; | |
7201 | ss << "We are recovering stretch mode buckets, only requiring " | |
7202 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
7203 | checks->add("RECOVERING_STRETCH_MODE", HEALTH_WARN, | |
7204 | ss.str(), 0); | |
7205 | } else if (degraded_stretch_mode) { | |
7206 | stringstream ss; | |
7207 | ss << "We are missing stretch mode buckets, only requiring " | |
7208 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
7209 | checks->add("DEGRADED_STRETCH_MODE", HEALTH_WARN, | |
7210 | ss.str(), 0); | |
7211 | } | |
7212 | } | |
aee94f69 TL |
7213 | // UNEQUAL_WEIGHT |
7214 | if (stretch_mode_enabled) { | |
7215 | vector<int> subtrees; | |
7216 | crush->get_subtree_of_type(stretch_mode_bucket, &subtrees); | |
7217 | if (subtrees.size() != 2) { | |
7218 | stringstream ss; | |
7219 | ss << "Stretch mode buckets != 2"; | |
7220 | checks->add("INCORRECT_NUM_BUCKETS_STRETCH_MODE", HEALTH_WARN, ss.str(), 0); | |
7221 | return; | |
7222 | } | |
7223 | int weight1 = crush->get_item_weight(subtrees[0]); | |
7224 | int weight2 = crush->get_item_weight(subtrees[1]); | |
7225 | stringstream ss; | |
7226 | if (weight1 != weight2) { | |
7227 | ss << "Stretch mode buckets have different weights!"; | |
7228 | checks->add("UNEVEN_WEIGHTS_STRETCH_MODE", HEALTH_WARN, ss.str(), 0); | |
7229 | } | |
7230 | } | |
224ce89b | 7231 | } |
35e4c445 FG |
7232 | |
7233 | int OSDMap::parse_osd_id_list(const vector<string>& ls, set<int> *out, | |
7234 | ostream *ss) const | |
7235 | { | |
7236 | out->clear(); | |
7237 | for (auto i = ls.begin(); i != ls.end(); ++i) { | |
7238 | if (i == ls.begin() && | |
7239 | (*i == "any" || *i == "all" || *i == "*")) { | |
7240 | get_all_osds(*out); | |
7241 | break; | |
7242 | } | |
1e59de90 | 7243 | long osd = ceph::common::parse_osd_id(i->c_str(), ss); |
35e4c445 FG |
7244 | if (osd < 0) { |
7245 | *ss << "invalid osd id '" << *i << "'"; | |
7246 | return -EINVAL; | |
7247 | } | |
7248 | out->insert(osd); | |
7249 | } | |
7250 | return 0; | |
7251 | } | |
11fdf7f2 TL |
7252 | |
7253 | void OSDMap::get_random_up_osds_by_subtree(int n, // whoami | |
7254 | string &subtree, | |
7255 | int limit, // how many | |
7256 | set<int> skip, | |
7257 | set<int> *want) const { | |
7258 | if (limit <= 0) | |
7259 | return; | |
7260 | int subtree_type = crush->get_type_id(subtree); | |
7261 | if (subtree_type < 1) | |
7262 | return; | |
7263 | vector<int> subtrees; | |
7264 | crush->get_subtree_of_type(subtree_type, &subtrees); | |
7265 | std::random_device rd; | |
7266 | std::default_random_engine rng{rd()}; | |
7267 | std::shuffle(subtrees.begin(), subtrees.end(), rng); | |
7268 | for (auto s : subtrees) { | |
7269 | if (limit <= 0) | |
7270 | break; | |
7271 | if (crush->subtree_contains(s, n)) | |
7272 | continue; | |
7273 | vector<int> osds; | |
7274 | crush->get_children_of_type(s, 0, &osds); | |
7275 | if (osds.empty()) | |
7276 | continue; | |
7277 | vector<int> up_osds; | |
7278 | for (auto o : osds) { | |
7279 | if (is_up(o) && !skip.count(o)) | |
7280 | up_osds.push_back(o); | |
7281 | } | |
7282 | if (up_osds.empty()) | |
7283 | continue; | |
7284 | auto it = up_osds.begin(); | |
7285 | std::advance(it, (n % up_osds.size())); | |
7286 | want->insert(*it); | |
7287 | --limit; | |
7288 | } | |
7289 | } | |
7290 | ||
7291 | float OSDMap::pool_raw_used_rate(int64_t poolid) const | |
7292 | { | |
7293 | const pg_pool_t *pool = get_pg_pool(poolid); | |
7294 | assert(pool != nullptr); | |
7295 | ||
7296 | switch (pool->get_type()) { | |
7297 | case pg_pool_t::TYPE_REPLICATED: | |
7298 | return pool->get_size(); | |
11fdf7f2 TL |
7299 | case pg_pool_t::TYPE_ERASURE: |
7300 | { | |
7301 | auto& ecp = | |
7302 | get_erasure_code_profile(pool->erasure_code_profile); | |
7303 | auto pm = ecp.find("m"); | |
7304 | auto pk = ecp.find("k"); | |
7305 | if (pm != ecp.end() && pk != ecp.end()) { | |
7306 | int k = atoi(pk->second.c_str()); | |
7307 | int m = atoi(pm->second.c_str()); | |
7308 | int mk = m + k; | |
7309 | ceph_assert(mk != 0); | |
7310 | ceph_assert(k != 0); | |
7311 | return (float)mk / k; | |
7312 | } else { | |
7313 | return 0.0; | |
7314 | } | |
7315 | } | |
7316 | break; | |
7317 | default: | |
7318 | ceph_abort_msg("unrecognized pool type"); | |
7319 | } | |
7320 | } | |
81eedcae TL |
7321 | |
7322 | unsigned OSDMap::get_osd_crush_node_flags(int osd) const | |
7323 | { | |
7324 | unsigned flags = 0; | |
7325 | if (!crush_node_flags.empty()) { | |
7326 | // the map will contain type -> name | |
7327 | std::map<std::string,std::string> ploc = crush->get_full_location(osd); | |
7328 | for (auto& i : ploc) { | |
7329 | int id = crush->get_item_id(i.second); | |
7330 | auto p = crush_node_flags.find(id); | |
7331 | if (p != crush_node_flags.end()) { | |
7332 | flags |= p->second; | |
7333 | } | |
7334 | } | |
7335 | } | |
7336 | return flags; | |
7337 | } | |
7338 | ||
7339 | unsigned OSDMap::get_crush_node_flags(int id) const | |
7340 | { | |
7341 | unsigned flags = 0; | |
7342 | auto it = crush_node_flags.find(id); | |
7343 | if (it != crush_node_flags.end()) | |
7344 | flags = it->second; | |
7345 | return flags; | |
7346 | } | |
7347 | ||
7348 | unsigned OSDMap::get_device_class_flags(int id) const | |
7349 | { | |
7350 | unsigned flags = 0; | |
7351 | auto it = device_class_flags.find(id); | |
7352 | if (it != device_class_flags.end()) | |
7353 | flags = it->second; | |
7354 | return flags; | |
7355 | } | |
20effc67 TL |
7356 | |
7357 | std::optional<std::string> OSDMap::pending_require_osd_release() const | |
7358 | { | |
7359 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_QUINCY) && | |
7360 | require_osd_release < ceph_release_t::quincy) { | |
7361 | return "quincy"; | |
7362 | } | |
7363 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_PACIFIC) && | |
7364 | require_osd_release < ceph_release_t::pacific) { | |
7365 | return "pacific"; | |
7366 | } | |
7367 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_OCTOPUS) && | |
7368 | require_osd_release < ceph_release_t::octopus) { | |
7369 | return "octopus"; | |
7370 | } | |
7371 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_NAUTILUS) && | |
7372 | require_osd_release < ceph_release_t::nautilus) { | |
7373 | return "nautilus"; | |
7374 | } | |
7375 | ||
7376 | return std::nullopt; | |
7377 | } |