]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com> | |
8 | * | |
9 | * Author: Loic Dachary <loic@dachary.org> | |
10 | * | |
11 | * This is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU Lesser General Public | |
13 | * License version 2.1, as published by the Free Software | |
14 | * Foundation. See file COPYING. | |
15 | * | |
16 | */ | |
17 | ||
11fdf7f2 TL |
18 | #include <algorithm> |
19 | #include <optional> | |
20 | #include <random> | |
21 | ||
224ce89b WB |
22 | #include <boost/algorithm/string.hpp> |
23 | ||
7c673cae | 24 | #include "OSDMap.h" |
7c673cae | 25 | #include "common/config.h" |
3efd9988 | 26 | #include "common/errno.h" |
7c673cae FG |
27 | #include "common/Formatter.h" |
28 | #include "common/TextTable.h" | |
29 | #include "include/ceph_features.h" | |
9f95a23c | 30 | #include "include/common_fwd.h" |
7c673cae FG |
31 | #include "include/str_map.h" |
32 | ||
33 | #include "common/code_environment.h" | |
224ce89b | 34 | #include "mon/health_check.h" |
7c673cae FG |
35 | |
36 | #include "crush/CrushTreeDumper.h" | |
37 | #include "common/Clock.h" | |
11fdf7f2 TL |
38 | #include "mon/PGMap.h" |
39 | ||
9f95a23c TL |
40 | using std::list; |
41 | using std::make_pair; | |
42 | using std::map; | |
43 | using std::multimap; | |
44 | using std::ostream; | |
45 | using std::ostringstream; | |
46 | using std::pair; | |
47 | using std::set; | |
48 | using std::string; | |
49 | using std::stringstream; | |
50 | using std::unordered_map; | |
51 | using std::vector; | |
52 | ||
53 | using ceph::decode; | |
54 | using ceph::encode; | |
55 | using ceph::Formatter; | |
56 | ||
7c673cae FG |
57 | #define dout_subsys ceph_subsys_osd |
58 | ||
59 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap, osdmap, osdmap); | |
60 | MEMPOOL_DEFINE_OBJECT_FACTORY(OSDMap::Incremental, osdmap_inc, osdmap); | |
61 | ||
62 | ||
63 | // ---------------------------------- | |
64 | // osd_info_t | |
65 | ||
66 | void osd_info_t::dump(Formatter *f) const | |
67 | { | |
68 | f->dump_int("last_clean_begin", last_clean_begin); | |
69 | f->dump_int("last_clean_end", last_clean_end); | |
70 | f->dump_int("up_from", up_from); | |
71 | f->dump_int("up_thru", up_thru); | |
72 | f->dump_int("down_at", down_at); | |
73 | f->dump_int("lost_at", lost_at); | |
74 | } | |
75 | ||
9f95a23c | 76 | void osd_info_t::encode(ceph::buffer::list& bl) const |
7c673cae | 77 | { |
11fdf7f2 | 78 | using ceph::encode; |
7c673cae | 79 | __u8 struct_v = 1; |
11fdf7f2 TL |
80 | encode(struct_v, bl); |
81 | encode(last_clean_begin, bl); | |
82 | encode(last_clean_end, bl); | |
83 | encode(up_from, bl); | |
84 | encode(up_thru, bl); | |
85 | encode(down_at, bl); | |
86 | encode(lost_at, bl); | |
7c673cae FG |
87 | } |
88 | ||
9f95a23c | 89 | void osd_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 90 | { |
11fdf7f2 | 91 | using ceph::decode; |
7c673cae | 92 | __u8 struct_v; |
11fdf7f2 TL |
93 | decode(struct_v, bl); |
94 | decode(last_clean_begin, bl); | |
95 | decode(last_clean_end, bl); | |
96 | decode(up_from, bl); | |
97 | decode(up_thru, bl); | |
98 | decode(down_at, bl); | |
99 | decode(lost_at, bl); | |
7c673cae FG |
100 | } |
101 | ||
102 | void osd_info_t::generate_test_instances(list<osd_info_t*>& o) | |
103 | { | |
104 | o.push_back(new osd_info_t); | |
105 | o.push_back(new osd_info_t); | |
106 | o.back()->last_clean_begin = 1; | |
107 | o.back()->last_clean_end = 2; | |
108 | o.back()->up_from = 30; | |
109 | o.back()->up_thru = 40; | |
110 | o.back()->down_at = 5; | |
111 | o.back()->lost_at = 6; | |
112 | } | |
113 | ||
114 | ostream& operator<<(ostream& out, const osd_info_t& info) | |
115 | { | |
116 | out << "up_from " << info.up_from | |
117 | << " up_thru " << info.up_thru | |
118 | << " down_at " << info.down_at | |
119 | << " last_clean_interval [" << info.last_clean_begin << "," << info.last_clean_end << ")"; | |
120 | if (info.lost_at) | |
121 | out << " lost_at " << info.lost_at; | |
122 | return out; | |
123 | } | |
124 | ||
125 | // ---------------------------------- | |
126 | // osd_xinfo_t | |
127 | ||
128 | void osd_xinfo_t::dump(Formatter *f) const | |
129 | { | |
130 | f->dump_stream("down_stamp") << down_stamp; | |
131 | f->dump_float("laggy_probability", laggy_probability); | |
132 | f->dump_int("laggy_interval", laggy_interval); | |
133 | f->dump_int("features", features); | |
134 | f->dump_unsigned("old_weight", old_weight); | |
9f95a23c TL |
135 | f->dump_stream("last_purged_snaps_scrub") << last_purged_snaps_scrub; |
136 | f->dump_int("dead_epoch", dead_epoch); | |
7c673cae FG |
137 | } |
138 | ||
9f95a23c | 139 | void osd_xinfo_t::encode(ceph::buffer::list& bl, uint64_t enc_features) const |
7c673cae | 140 | { |
9f95a23c TL |
141 | uint8_t v = 4; |
142 | if (!HAVE_FEATURE(enc_features, SERVER_OCTOPUS)) { | |
143 | v = 3; | |
144 | } | |
145 | ENCODE_START(v, 1, bl); | |
11fdf7f2 | 146 | encode(down_stamp, bl); |
f67539c2 | 147 | __u32 lp = laggy_probability * float(0xfffffffful); |
11fdf7f2 TL |
148 | encode(lp, bl); |
149 | encode(laggy_interval, bl); | |
150 | encode(features, bl); | |
151 | encode(old_weight, bl); | |
9f95a23c TL |
152 | if (v >= 4) { |
153 | encode(last_purged_snaps_scrub, bl); | |
154 | encode(dead_epoch, bl); | |
155 | } | |
7c673cae FG |
156 | ENCODE_FINISH(bl); |
157 | } | |
158 | ||
9f95a23c | 159 | void osd_xinfo_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 160 | { |
9f95a23c | 161 | DECODE_START(4, bl); |
11fdf7f2 | 162 | decode(down_stamp, bl); |
7c673cae | 163 | __u32 lp; |
11fdf7f2 | 164 | decode(lp, bl); |
7c673cae | 165 | laggy_probability = (float)lp / (float)0xffffffff; |
11fdf7f2 | 166 | decode(laggy_interval, bl); |
7c673cae | 167 | if (struct_v >= 2) |
11fdf7f2 | 168 | decode(features, bl); |
7c673cae FG |
169 | else |
170 | features = 0; | |
171 | if (struct_v >= 3) | |
11fdf7f2 | 172 | decode(old_weight, bl); |
7c673cae FG |
173 | else |
174 | old_weight = 0; | |
9f95a23c TL |
175 | if (struct_v >= 4) { |
176 | decode(last_purged_snaps_scrub, bl); | |
177 | decode(dead_epoch, bl); | |
178 | } else { | |
179 | dead_epoch = 0; | |
180 | } | |
7c673cae FG |
181 | DECODE_FINISH(bl); |
182 | } | |
183 | ||
184 | void osd_xinfo_t::generate_test_instances(list<osd_xinfo_t*>& o) | |
185 | { | |
186 | o.push_back(new osd_xinfo_t); | |
187 | o.push_back(new osd_xinfo_t); | |
188 | o.back()->down_stamp = utime_t(2, 3); | |
189 | o.back()->laggy_probability = .123; | |
190 | o.back()->laggy_interval = 123456; | |
191 | o.back()->old_weight = 0x7fff; | |
192 | } | |
193 | ||
194 | ostream& operator<<(ostream& out, const osd_xinfo_t& xi) | |
195 | { | |
196 | return out << "down_stamp " << xi.down_stamp | |
197 | << " laggy_probability " << xi.laggy_probability | |
198 | << " laggy_interval " << xi.laggy_interval | |
9f95a23c TL |
199 | << " old_weight " << xi.old_weight |
200 | << " last_purged_snaps_scrub " << xi.last_purged_snaps_scrub | |
201 | << " dead_epoch " << xi.dead_epoch; | |
7c673cae FG |
202 | } |
203 | ||
204 | // ---------------------------------- | |
205 | // OSDMap::Incremental | |
206 | ||
207 | int OSDMap::Incremental::get_net_marked_out(const OSDMap *previous) const | |
208 | { | |
209 | int n = 0; | |
210 | for (auto &weight : new_weight) { | |
211 | if (weight.second == CEPH_OSD_OUT && !previous->is_out(weight.first)) | |
212 | n++; // marked out | |
213 | else if (weight.second != CEPH_OSD_OUT && previous->is_out(weight.first)) | |
214 | n--; // marked in | |
215 | } | |
216 | return n; | |
217 | } | |
218 | ||
219 | int OSDMap::Incremental::get_net_marked_down(const OSDMap *previous) const | |
220 | { | |
221 | int n = 0; | |
222 | for (auto &state : new_state) { // | |
223 | if (state.second & CEPH_OSD_UP) { | |
224 | if (previous->is_up(state.first)) | |
225 | n++; // marked down | |
226 | else | |
227 | n--; // marked up | |
228 | } | |
229 | } | |
230 | return n; | |
231 | } | |
232 | ||
233 | int OSDMap::Incremental::identify_osd(uuid_d u) const | |
234 | { | |
235 | for (auto &uuid : new_uuid) | |
236 | if (uuid.second == u) | |
237 | return uuid.first; | |
238 | return -1; | |
239 | } | |
240 | ||
f67539c2 TL |
241 | int OSDMap::Incremental::propagate_base_properties_to_tiers(CephContext *cct, |
242 | const OSDMap& osdmap) | |
7c673cae | 243 | { |
11fdf7f2 | 244 | ceph_assert(epoch == osdmap.get_epoch() + 1); |
7c673cae FG |
245 | |
246 | for (auto &new_pool : new_pools) { | |
247 | if (!new_pool.second.tiers.empty()) { | |
248 | pg_pool_t& base = new_pool.second; | |
249 | ||
11fdf7f2 TL |
250 | auto new_rem_it = new_removed_snaps.find(new_pool.first); |
251 | ||
7c673cae FG |
252 | for (const auto &tier_pool : base.tiers) { |
253 | const auto &r = new_pools.find(tier_pool); | |
254 | pg_pool_t *tier = 0; | |
255 | if (r == new_pools.end()) { | |
256 | const pg_pool_t *orig = osdmap.get_pg_pool(tier_pool); | |
257 | if (!orig) { | |
258 | lderr(cct) << __func__ << " no pool " << tier_pool << dendl; | |
259 | return -EIO; | |
260 | } | |
261 | tier = get_new_pool(tier_pool, orig); | |
262 | } else { | |
263 | tier = &r->second; | |
264 | } | |
265 | if (tier->tier_of != new_pool.first) { | |
266 | lderr(cct) << __func__ << " " << r->first << " tier_of != " << new_pool.first << dendl; | |
267 | return -EIO; | |
268 | } | |
269 | ||
270 | ldout(cct, 10) << __func__ << " from " << new_pool.first << " to " | |
271 | << tier_pool << dendl; | |
272 | tier->snap_seq = base.snap_seq; | |
273 | tier->snap_epoch = base.snap_epoch; | |
274 | tier->snaps = base.snaps; | |
275 | tier->removed_snaps = base.removed_snaps; | |
11fdf7f2 TL |
276 | tier->flags |= base.flags & (pg_pool_t::FLAG_SELFMANAGED_SNAPS| |
277 | pg_pool_t::FLAG_POOL_SNAPS); | |
278 | ||
279 | if (new_rem_it != new_removed_snaps.end()) { | |
280 | new_removed_snaps[tier_pool] = new_rem_it->second; | |
281 | } | |
f67539c2 TL |
282 | |
283 | tier->application_metadata = base.application_metadata; | |
7c673cae FG |
284 | } |
285 | } | |
286 | } | |
287 | return 0; | |
288 | } | |
289 | ||
28e407b8 AA |
290 | // ---------------------------------- |
291 | // OSDMap | |
7c673cae FG |
292 | |
293 | bool OSDMap::subtree_is_down(int id, set<int> *down_cache) const | |
294 | { | |
295 | if (id >= 0) | |
296 | return is_down(id); | |
297 | ||
298 | if (down_cache && | |
299 | down_cache->count(id)) { | |
300 | return true; | |
301 | } | |
302 | ||
303 | list<int> children; | |
304 | crush->get_children(id, &children); | |
305 | for (const auto &child : children) { | |
306 | if (!subtree_is_down(child, down_cache)) { | |
307 | return false; | |
308 | } | |
309 | } | |
310 | if (down_cache) { | |
311 | down_cache->insert(id); | |
312 | } | |
313 | return true; | |
314 | } | |
315 | ||
316 | bool OSDMap::containing_subtree_is_down(CephContext *cct, int id, int subtree_type, set<int> *down_cache) const | |
317 | { | |
318 | // use a stack-local down_cache if we didn't get one from the | |
319 | // caller. then at least this particular call will avoid duplicated | |
320 | // work. | |
321 | set<int> local_down_cache; | |
322 | if (!down_cache) { | |
323 | down_cache = &local_down_cache; | |
324 | } | |
325 | ||
326 | int current = id; | |
327 | while (true) { | |
328 | int type; | |
329 | if (current >= 0) { | |
330 | type = 0; | |
331 | } else { | |
332 | type = crush->get_bucket_type(current); | |
333 | } | |
11fdf7f2 | 334 | ceph_assert(type >= 0); |
7c673cae FG |
335 | |
336 | if (!subtree_is_down(current, down_cache)) { | |
337 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = false" << dendl; | |
338 | return false; | |
339 | } | |
340 | ||
341 | // is this a big enough subtree to be marked as down? | |
342 | if (type >= subtree_type) { | |
343 | ldout(cct, 30) << "containing_subtree_is_down(" << id << ") = true ... " << type << " >= " << subtree_type << dendl; | |
344 | return true; | |
345 | } | |
346 | ||
347 | int r = crush->get_immediate_parent_id(current, ¤t); | |
348 | if (r < 0) { | |
349 | return false; | |
350 | } | |
351 | } | |
352 | } | |
353 | ||
224ce89b WB |
354 | bool OSDMap::subtree_type_is_down( |
355 | CephContext *cct, | |
356 | int id, | |
357 | int subtree_type, | |
358 | set<int> *down_in_osds, | |
359 | set<int> *up_in_osds, | |
360 | set<int> *subtree_up, | |
361 | unordered_map<int, set<int> > *subtree_type_down) const | |
31f18b77 FG |
362 | { |
363 | if (id >= 0) { | |
364 | bool is_down_ret = is_down(id); | |
365 | if (!is_out(id)) { | |
366 | if (is_down_ret) { | |
367 | down_in_osds->insert(id); | |
368 | } else { | |
369 | up_in_osds->insert(id); | |
370 | } | |
371 | } | |
372 | return is_down_ret; | |
373 | } | |
374 | ||
375 | if (subtree_type_down && | |
376 | (*subtree_type_down)[subtree_type].count(id)) { | |
377 | return true; | |
378 | } | |
379 | ||
380 | list<int> children; | |
381 | crush->get_children(id, &children); | |
382 | for (const auto &child : children) { | |
224ce89b WB |
383 | if (!subtree_type_is_down( |
384 | cct, child, crush->get_bucket_type(child), | |
385 | down_in_osds, up_in_osds, subtree_up, subtree_type_down)) { | |
31f18b77 FG |
386 | subtree_up->insert(id); |
387 | return false; | |
388 | } | |
389 | } | |
390 | if (subtree_type_down) { | |
391 | (*subtree_type_down)[subtree_type].insert(id); | |
392 | } | |
393 | return true; | |
394 | } | |
395 | ||
9f95a23c | 396 | void OSDMap::Incremental::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 397 | { |
11fdf7f2 | 398 | using ceph::encode; |
7c673cae | 399 | __u16 v = 5; |
11fdf7f2 TL |
400 | encode(v, bl); |
401 | encode(fsid, bl); | |
402 | encode(epoch, bl); | |
403 | encode(modified, bl); | |
7c673cae | 404 | int32_t new_t = new_pool_max; |
11fdf7f2 TL |
405 | encode(new_t, bl); |
406 | encode(new_flags, bl); | |
407 | encode(fullmap, bl); | |
408 | encode(crush, bl); | |
7c673cae | 409 | |
11fdf7f2 TL |
410 | encode(new_max_osd, bl); |
411 | // for encode(new_pools, bl); | |
7c673cae | 412 | __u32 n = new_pools.size(); |
11fdf7f2 | 413 | encode(n, bl); |
7c673cae FG |
414 | for (const auto &new_pool : new_pools) { |
415 | n = new_pool.first; | |
11fdf7f2 TL |
416 | encode(n, bl); |
417 | encode(new_pool.second, bl, 0); | |
7c673cae | 418 | } |
11fdf7f2 | 419 | // for encode(new_pool_names, bl); |
7c673cae | 420 | n = new_pool_names.size(); |
11fdf7f2 | 421 | encode(n, bl); |
7c673cae FG |
422 | |
423 | for (const auto &new_pool_name : new_pool_names) { | |
424 | n = new_pool_name.first; | |
11fdf7f2 TL |
425 | encode(n, bl); |
426 | encode(new_pool_name.second, bl); | |
7c673cae | 427 | } |
11fdf7f2 | 428 | // for encode(old_pools, bl); |
7c673cae | 429 | n = old_pools.size(); |
11fdf7f2 | 430 | encode(n, bl); |
7c673cae FG |
431 | for (auto &old_pool : old_pools) { |
432 | n = old_pool; | |
11fdf7f2 | 433 | encode(n, bl); |
7c673cae | 434 | } |
11fdf7f2 | 435 | encode(new_up_client, bl, 0); |
31f18b77 FG |
436 | { |
437 | // legacy is map<int32_t,uint8_t> | |
9f95a23c | 438 | map<int32_t, uint8_t> os; |
31f18b77 | 439 | for (auto p : new_state) { |
9f95a23c TL |
440 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
441 | // that an old client could not understand. | |
442 | // skip those! | |
443 | uint8_t s = p.second; | |
444 | if (p.second != 0 && s == 0) | |
445 | continue; | |
446 | os[p.first] = s; | |
447 | } | |
448 | uint32_t n = os.size(); | |
449 | encode(n, bl); | |
450 | for (auto p : os) { | |
11fdf7f2 | 451 | encode(p.first, bl); |
9f95a23c | 452 | encode(p.second, bl); |
31f18b77 FG |
453 | } |
454 | } | |
11fdf7f2 TL |
455 | encode(new_weight, bl); |
456 | // for encode(new_pg_temp, bl); | |
7c673cae | 457 | n = new_pg_temp.size(); |
11fdf7f2 | 458 | encode(n, bl); |
7c673cae FG |
459 | |
460 | for (const auto &pg_temp : new_pg_temp) { | |
461 | old_pg_t opg = pg_temp.first.get_old_pg(); | |
11fdf7f2 TL |
462 | encode(opg, bl); |
463 | encode(pg_temp.second, bl); | |
7c673cae FG |
464 | } |
465 | } | |
466 | ||
9f95a23c | 467 | void OSDMap::Incremental::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 468 | { |
11fdf7f2 | 469 | using ceph::encode; |
7c673cae FG |
470 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
471 | encode_client_old(bl); | |
472 | return; | |
473 | } | |
474 | ||
475 | // base | |
476 | __u16 v = 6; | |
11fdf7f2 TL |
477 | encode(v, bl); |
478 | encode(fsid, bl); | |
479 | encode(epoch, bl); | |
480 | encode(modified, bl); | |
481 | encode(new_pool_max, bl); | |
482 | encode(new_flags, bl); | |
483 | encode(fullmap, bl); | |
484 | encode(crush, bl); | |
485 | ||
486 | encode(new_max_osd, bl); | |
487 | encode(new_pools, bl, features); | |
488 | encode(new_pool_names, bl); | |
489 | encode(old_pools, bl); | |
490 | encode(new_up_client, bl, features); | |
31f18b77 | 491 | { |
9f95a23c | 492 | map<int32_t, uint8_t> os; |
31f18b77 | 493 | for (auto p : new_state) { |
9f95a23c TL |
494 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
495 | // that an old client could not understand. | |
496 | // skip those! | |
497 | uint8_t s = p.second; | |
498 | if (p.second != 0 && s == 0) | |
499 | continue; | |
500 | os[p.first] = s; | |
501 | } | |
502 | uint32_t n = os.size(); | |
503 | encode(n, bl); | |
504 | for (auto p : os) { | |
11fdf7f2 | 505 | encode(p.first, bl); |
9f95a23c | 506 | encode(p.second, bl); |
31f18b77 FG |
507 | } |
508 | } | |
11fdf7f2 TL |
509 | encode(new_weight, bl); |
510 | encode(new_pg_temp, bl); | |
7c673cae FG |
511 | |
512 | // extended | |
513 | __u16 ev = 10; | |
11fdf7f2 TL |
514 | encode(ev, bl); |
515 | encode(new_hb_back_up, bl, features); | |
516 | encode(new_up_thru, bl); | |
517 | encode(new_last_clean_interval, bl); | |
518 | encode(new_lost, bl); | |
f67539c2 TL |
519 | encode(new_blocklist, bl, features); |
520 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
521 | encode(new_up_cluster, bl, features); |
522 | encode(cluster_snapshot, bl); | |
523 | encode(new_uuid, bl); | |
9f95a23c | 524 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
525 | encode(new_hb_front_up, bl, features); |
526 | } | |
527 | ||
528 | template<class T> | |
9f95a23c | 529 | static void encode_addrvec_map_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
530 | { |
531 | uint32_t n = m.size(); | |
532 | encode(n, bl); | |
533 | for (auto& i : m) { | |
534 | encode(i.first, bl); | |
535 | encode(i.second.legacy_addr(), bl, f); | |
536 | } | |
537 | } | |
538 | ||
539 | template<class T> | |
9f95a23c | 540 | static void encode_addrvec_pvec_as_addr(const T& m, ceph::buffer::list& bl, uint64_t f) |
11fdf7f2 TL |
541 | { |
542 | uint32_t n = m.size(); | |
543 | encode(n, bl); | |
544 | for (auto& i : m) { | |
545 | if (i) { | |
546 | encode(i->legacy_addr(), bl, f); | |
547 | } else { | |
548 | encode(entity_addr_t(), bl, f); | |
549 | } | |
550 | } | |
7c673cae FG |
551 | } |
552 | ||
11fdf7f2 TL |
553 | /* for a description of osdmap incremental versions, and when they were |
554 | * introduced, please refer to | |
555 | * doc/dev/osd_internals/osdmap_versions.txt | |
556 | */ | |
9f95a23c | 557 | void OSDMap::Incremental::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 558 | { |
11fdf7f2 | 559 | using ceph::encode; |
7c673cae FG |
560 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
561 | encode_classic(bl, features); | |
562 | return; | |
563 | } | |
564 | ||
565 | // only a select set of callers should *ever* be encoding new | |
566 | // OSDMaps. others should be passing around the canonical encoded | |
567 | // buffers from on high. select out those callers by passing in an | |
568 | // "impossible" feature bit. | |
11fdf7f2 | 569 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
570 | features &= ~CEPH_FEATURE_RESERVED; |
571 | ||
572 | size_t start_offset = bl.length(); | |
573 | size_t tail_offset; | |
11fdf7f2 | 574 | size_t crc_offset; |
9f95a23c | 575 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
576 | |
577 | // meta-encoding: how we include client-used and osd-specific data | |
578 | ENCODE_START(8, 7, bl); | |
579 | ||
580 | { | |
11fdf7f2 | 581 | uint8_t v = 8; |
7c673cae FG |
582 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
583 | v = 3; | |
11fdf7f2 TL |
584 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
585 | v = 5; | |
586 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
587 | v = 6; | |
7c673cae FG |
588 | } |
589 | ENCODE_START(v, 1, bl); // client-usable data | |
11fdf7f2 TL |
590 | encode(fsid, bl); |
591 | encode(epoch, bl); | |
592 | encode(modified, bl); | |
593 | encode(new_pool_max, bl); | |
594 | encode(new_flags, bl); | |
595 | encode(fullmap, bl); | |
596 | encode(crush, bl); | |
597 | ||
598 | encode(new_max_osd, bl); | |
599 | encode(new_pools, bl, features); | |
600 | encode(new_pool_names, bl); | |
601 | encode(old_pools, bl); | |
602 | if (v >= 7) { | |
603 | encode(new_up_client, bl, features); | |
604 | } else { | |
605 | encode_addrvec_map_as_addr(new_up_client, bl, features); | |
606 | } | |
31f18b77 | 607 | if (v >= 5) { |
11fdf7f2 | 608 | encode(new_state, bl); |
31f18b77 | 609 | } else { |
9f95a23c | 610 | map<int32_t, uint8_t> os; |
31f18b77 | 611 | for (auto p : new_state) { |
9f95a23c TL |
612 | // new_state may only inculde some new flags(e.g., CEPH_OSD_NOOUT) |
613 | // that an old client could not understand. | |
614 | // skip those! | |
615 | uint8_t s = p.second; | |
616 | if (p.second != 0 && s == 0) | |
617 | continue; | |
618 | os[p.first] = s; | |
619 | } | |
620 | uint32_t n = os.size(); | |
621 | encode(n, bl); | |
622 | for (auto p : os) { | |
623 | encode(p.first, bl); | |
624 | encode(p.second, bl); | |
31f18b77 FG |
625 | } |
626 | } | |
11fdf7f2 TL |
627 | encode(new_weight, bl); |
628 | encode(new_pg_temp, bl); | |
629 | encode(new_primary_temp, bl); | |
630 | encode(new_primary_affinity, bl); | |
631 | encode(new_erasure_code_profiles, bl); | |
632 | encode(old_erasure_code_profiles, bl); | |
7c673cae | 633 | if (v >= 4) { |
11fdf7f2 TL |
634 | encode(new_pg_upmap, bl); |
635 | encode(old_pg_upmap, bl); | |
636 | encode(new_pg_upmap_items, bl); | |
637 | encode(old_pg_upmap_items, bl); | |
638 | } | |
639 | if (v >= 6) { | |
640 | encode(new_removed_snaps, bl); | |
641 | encode(new_purged_snaps, bl); | |
642 | } | |
643 | if (v >= 8) { | |
644 | encode(new_last_up_change, bl); | |
645 | encode(new_last_in_change, bl); | |
7c673cae FG |
646 | } |
647 | ENCODE_FINISH(bl); // client-usable data | |
648 | } | |
649 | ||
650 | { | |
33c7a0ef | 651 | uint8_t target_v = 9; // if bumping this, be aware of range_blocklist 11 |
7c673cae FG |
652 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
653 | target_v = 2; | |
11fdf7f2 TL |
654 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { |
655 | target_v = 6; | |
7c673cae | 656 | } |
f67539c2 | 657 | if (change_stretch_mode) { |
f67539c2 TL |
658 | target_v = std::max((uint8_t)10, target_v); |
659 | } | |
33c7a0ef TL |
660 | if (!new_range_blocklist.empty() || |
661 | !old_range_blocklist.empty()) { | |
662 | target_v = std::max((uint8_t)11, target_v); | |
663 | } | |
7c673cae | 664 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
665 | if (target_v < 7) { |
666 | encode_addrvec_map_as_addr(new_hb_back_up, bl, features); | |
667 | } else { | |
668 | encode(new_hb_back_up, bl, features); | |
669 | } | |
670 | encode(new_up_thru, bl); | |
671 | encode(new_last_clean_interval, bl); | |
672 | encode(new_lost, bl); | |
f67539c2 TL |
673 | encode(new_blocklist, bl, features); |
674 | encode(old_blocklist, bl, features); | |
11fdf7f2 TL |
675 | if (target_v < 7) { |
676 | encode_addrvec_map_as_addr(new_up_cluster, bl, features); | |
677 | } else { | |
678 | encode(new_up_cluster, bl, features); | |
679 | } | |
680 | encode(cluster_snapshot, bl); | |
681 | encode(new_uuid, bl); | |
9f95a23c | 682 | encode(new_xinfo, bl, features); |
11fdf7f2 TL |
683 | if (target_v < 7) { |
684 | encode_addrvec_map_as_addr(new_hb_front_up, bl, features); | |
685 | } else { | |
686 | encode(new_hb_front_up, bl, features); | |
687 | } | |
688 | encode(features, bl); // NOTE: features arg, not the member | |
7c673cae | 689 | if (target_v >= 3) { |
11fdf7f2 TL |
690 | encode(new_nearfull_ratio, bl); |
691 | encode(new_full_ratio, bl); | |
692 | encode(new_backfillfull_ratio, bl); | |
31f18b77 FG |
693 | } |
694 | // 5 was string-based new_require_min_compat_client | |
695 | if (target_v >= 6) { | |
11fdf7f2 TL |
696 | encode(new_require_min_compat_client, bl); |
697 | encode(new_require_osd_release, bl); | |
7c673cae | 698 | } |
81eedcae TL |
699 | if (target_v >= 8) { |
700 | encode(new_crush_node_flags, bl); | |
701 | } | |
702 | if (target_v >= 9) { | |
703 | encode(new_device_class_flags, bl); | |
704 | } | |
f67539c2 TL |
705 | if (target_v >= 10) { |
706 | encode(change_stretch_mode, bl); | |
707 | encode(new_stretch_bucket_count, bl); | |
708 | encode(new_degraded_stretch_mode, bl); | |
709 | encode(new_recovering_stretch_mode, bl); | |
710 | encode(new_stretch_mode_bucket, bl); | |
711 | encode(stretch_mode_enabled, bl); | |
712 | } | |
33c7a0ef TL |
713 | if (target_v >= 11) { |
714 | encode(new_range_blocklist, bl, features); | |
715 | encode(old_range_blocklist, bl, features); | |
716 | } | |
7c673cae FG |
717 | ENCODE_FINISH(bl); // osd-only data |
718 | } | |
719 | ||
11fdf7f2 TL |
720 | crc_offset = bl.length(); |
721 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
722 | tail_offset = bl.length(); |
723 | ||
11fdf7f2 | 724 | encode(full_crc, bl); |
7c673cae FG |
725 | |
726 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
727 | ||
728 | // fill in crc | |
9f95a23c | 729 | ceph::buffer::list front; |
11fdf7f2 | 730 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae | 731 | inc_crc = front.crc32c(-1); |
9f95a23c | 732 | ceph::buffer::list tail; |
7c673cae FG |
733 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
734 | inc_crc = tail.crc32c(inc_crc); | |
735 | ceph_le32 crc_le; | |
736 | crc_le = inc_crc; | |
11fdf7f2 | 737 | crc_filler->copy_in(4u, (char*)&crc_le); |
7c673cae FG |
738 | have_crc = true; |
739 | } | |
740 | ||
9f95a23c | 741 | void OSDMap::Incremental::decode_classic(ceph::buffer::list::const_iterator &p) |
7c673cae | 742 | { |
11fdf7f2 | 743 | using ceph::decode; |
7c673cae FG |
744 | __u32 n, t; |
745 | // base | |
746 | __u16 v; | |
11fdf7f2 TL |
747 | decode(v, p); |
748 | decode(fsid, p); | |
749 | decode(epoch, p); | |
750 | decode(modified, p); | |
7c673cae | 751 | if (v == 4 || v == 5) { |
11fdf7f2 | 752 | decode(n, p); |
7c673cae FG |
753 | new_pool_max = n; |
754 | } else if (v >= 6) | |
11fdf7f2 TL |
755 | decode(new_pool_max, p); |
756 | decode(new_flags, p); | |
757 | decode(fullmap, p); | |
758 | decode(crush, p); | |
7c673cae | 759 | |
11fdf7f2 | 760 | decode(new_max_osd, p); |
7c673cae FG |
761 | if (v < 6) { |
762 | new_pools.clear(); | |
11fdf7f2 | 763 | decode(n, p); |
7c673cae | 764 | while (n--) { |
11fdf7f2 TL |
765 | decode(t, p); |
766 | decode(new_pools[t], p); | |
7c673cae FG |
767 | } |
768 | } else { | |
11fdf7f2 | 769 | decode(new_pools, p); |
7c673cae FG |
770 | } |
771 | if (v == 5) { | |
772 | new_pool_names.clear(); | |
11fdf7f2 | 773 | decode(n, p); |
7c673cae | 774 | while (n--) { |
11fdf7f2 TL |
775 | decode(t, p); |
776 | decode(new_pool_names[t], p); | |
7c673cae FG |
777 | } |
778 | } else if (v >= 6) { | |
11fdf7f2 | 779 | decode(new_pool_names, p); |
7c673cae FG |
780 | } |
781 | if (v < 6) { | |
782 | old_pools.clear(); | |
11fdf7f2 | 783 | decode(n, p); |
7c673cae | 784 | while (n--) { |
11fdf7f2 | 785 | decode(t, p); |
7c673cae FG |
786 | old_pools.insert(t); |
787 | } | |
788 | } else { | |
11fdf7f2 | 789 | decode(old_pools, p); |
7c673cae | 790 | } |
11fdf7f2 | 791 | decode(new_up_client, p); |
31f18b77 FG |
792 | { |
793 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 794 | decode(ns, p); |
31f18b77 FG |
795 | for (auto q : ns) { |
796 | new_state[q.first] = q.second; | |
797 | } | |
798 | } | |
11fdf7f2 | 799 | decode(new_weight, p); |
7c673cae FG |
800 | |
801 | if (v < 6) { | |
802 | new_pg_temp.clear(); | |
11fdf7f2 | 803 | decode(n, p); |
7c673cae FG |
804 | while (n--) { |
805 | old_pg_t opg; | |
9f95a23c | 806 | ceph::decode_raw(opg, p); |
11fdf7f2 | 807 | decode(new_pg_temp[pg_t(opg)], p); |
7c673cae FG |
808 | } |
809 | } else { | |
11fdf7f2 | 810 | decode(new_pg_temp, p); |
7c673cae FG |
811 | } |
812 | ||
813 | // decode short map, too. | |
814 | if (v == 5 && p.end()) | |
815 | return; | |
816 | ||
817 | // extended | |
818 | __u16 ev = 0; | |
819 | if (v >= 5) | |
11fdf7f2 TL |
820 | decode(ev, p); |
821 | decode(new_hb_back_up, p); | |
7c673cae | 822 | if (v < 5) |
11fdf7f2 TL |
823 | decode(new_pool_names, p); |
824 | decode(new_up_thru, p); | |
825 | decode(new_last_clean_interval, p); | |
826 | decode(new_lost, p); | |
f67539c2 TL |
827 | decode(new_blocklist, p); |
828 | decode(old_blocklist, p); | |
7c673cae | 829 | if (ev >= 6) |
11fdf7f2 | 830 | decode(new_up_cluster, p); |
7c673cae | 831 | if (ev >= 7) |
11fdf7f2 | 832 | decode(cluster_snapshot, p); |
7c673cae | 833 | if (ev >= 8) |
11fdf7f2 | 834 | decode(new_uuid, p); |
7c673cae | 835 | if (ev >= 9) |
11fdf7f2 | 836 | decode(new_xinfo, p); |
7c673cae | 837 | if (ev >= 10) |
11fdf7f2 | 838 | decode(new_hb_front_up, p); |
7c673cae FG |
839 | } |
840 | ||
11fdf7f2 TL |
841 | /* for a description of osdmap incremental versions, and when they were |
842 | * introduced, please refer to | |
843 | * doc/dev/osd_internals/osdmap_versions.txt | |
844 | */ | |
9f95a23c | 845 | void OSDMap::Incremental::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 846 | { |
11fdf7f2 | 847 | using ceph::decode; |
7c673cae FG |
848 | /** |
849 | * Older encodings of the Incremental had a single struct_v which | |
850 | * covered the whole encoding, and was prior to our modern | |
851 | * stuff which includes a compatv and a size. So if we see | |
852 | * a struct_v < 7, we must rewind to the beginning and use our | |
853 | * classic decoder. | |
854 | */ | |
855 | size_t start_offset = bl.get_off(); | |
856 | size_t tail_offset = 0; | |
9f95a23c | 857 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
858 | |
859 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
860 | if (struct_v < 7) { | |
11fdf7f2 | 861 | bl.seek(start_offset); |
7c673cae FG |
862 | decode_classic(bl); |
863 | encode_features = 0; | |
864 | if (struct_v >= 6) | |
865 | encode_features = CEPH_FEATURE_PGID64; | |
866 | else | |
867 | encode_features = 0; | |
868 | return; | |
869 | } | |
870 | { | |
11fdf7f2 TL |
871 | DECODE_START(8, bl); // client-usable data |
872 | decode(fsid, bl); | |
873 | decode(epoch, bl); | |
874 | decode(modified, bl); | |
875 | decode(new_pool_max, bl); | |
876 | decode(new_flags, bl); | |
877 | decode(fullmap, bl); | |
878 | decode(crush, bl); | |
879 | ||
880 | decode(new_max_osd, bl); | |
881 | decode(new_pools, bl); | |
882 | decode(new_pool_names, bl); | |
883 | decode(old_pools, bl); | |
884 | decode(new_up_client, bl); | |
31f18b77 | 885 | if (struct_v >= 5) { |
11fdf7f2 | 886 | decode(new_state, bl); |
31f18b77 FG |
887 | } else { |
888 | map<int32_t,uint8_t> ns; | |
11fdf7f2 | 889 | decode(ns, bl); |
31f18b77 FG |
890 | for (auto q : ns) { |
891 | new_state[q.first] = q.second; | |
892 | } | |
893 | } | |
11fdf7f2 TL |
894 | decode(new_weight, bl); |
895 | decode(new_pg_temp, bl); | |
896 | decode(new_primary_temp, bl); | |
7c673cae | 897 | if (struct_v >= 2) |
11fdf7f2 | 898 | decode(new_primary_affinity, bl); |
7c673cae FG |
899 | else |
900 | new_primary_affinity.clear(); | |
901 | if (struct_v >= 3) { | |
11fdf7f2 TL |
902 | decode(new_erasure_code_profiles, bl); |
903 | decode(old_erasure_code_profiles, bl); | |
7c673cae FG |
904 | } else { |
905 | new_erasure_code_profiles.clear(); | |
906 | old_erasure_code_profiles.clear(); | |
907 | } | |
908 | if (struct_v >= 4) { | |
11fdf7f2 TL |
909 | decode(new_pg_upmap, bl); |
910 | decode(old_pg_upmap, bl); | |
911 | decode(new_pg_upmap_items, bl); | |
912 | decode(old_pg_upmap_items, bl); | |
913 | } | |
914 | if (struct_v >= 6) { | |
915 | decode(new_removed_snaps, bl); | |
916 | decode(new_purged_snaps, bl); | |
917 | } | |
918 | if (struct_v >= 8) { | |
919 | decode(new_last_up_change, bl); | |
920 | decode(new_last_in_change, bl); | |
7c673cae FG |
921 | } |
922 | DECODE_FINISH(bl); // client-usable data | |
923 | } | |
924 | ||
925 | { | |
f67539c2 | 926 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
927 | decode(new_hb_back_up, bl); |
928 | decode(new_up_thru, bl); | |
929 | decode(new_last_clean_interval, bl); | |
930 | decode(new_lost, bl); | |
f67539c2 TL |
931 | decode(new_blocklist, bl); |
932 | decode(old_blocklist, bl); | |
11fdf7f2 TL |
933 | decode(new_up_cluster, bl); |
934 | decode(cluster_snapshot, bl); | |
935 | decode(new_uuid, bl); | |
936 | decode(new_xinfo, bl); | |
937 | decode(new_hb_front_up, bl); | |
7c673cae | 938 | if (struct_v >= 2) |
11fdf7f2 | 939 | decode(encode_features, bl); |
7c673cae FG |
940 | else |
941 | encode_features = CEPH_FEATURE_PGID64 | CEPH_FEATURE_OSDMAP_ENC; | |
942 | if (struct_v >= 3) { | |
11fdf7f2 TL |
943 | decode(new_nearfull_ratio, bl); |
944 | decode(new_full_ratio, bl); | |
7c673cae FG |
945 | } else { |
946 | new_nearfull_ratio = -1; | |
947 | new_full_ratio = -1; | |
948 | } | |
949 | if (struct_v >= 4) { | |
11fdf7f2 | 950 | decode(new_backfillfull_ratio, bl); |
7c673cae FG |
951 | } else { |
952 | new_backfillfull_ratio = -1; | |
953 | } | |
31f18b77 FG |
954 | if (struct_v == 5) { |
955 | string r; | |
11fdf7f2 | 956 | decode(r, bl); |
31f18b77 | 957 | if (r.length()) { |
9f95a23c | 958 | new_require_min_compat_client = ceph_release_from_name(r); |
31f18b77 FG |
959 | } |
960 | } | |
961 | if (struct_v >= 6) { | |
11fdf7f2 TL |
962 | decode(new_require_min_compat_client, bl); |
963 | decode(new_require_osd_release, bl); | |
31f18b77 FG |
964 | } else { |
965 | if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_LUMINOUS)) { | |
966 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 967 | new_require_osd_release = ceph_release_t::luminous; |
31f18b77 FG |
968 | new_flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
969 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_KRAKEN)) { | |
9f95a23c | 970 | new_require_osd_release = ceph_release_t::kraken; |
31f18b77 | 971 | } else if (new_flags >= 0 && (new_flags & CEPH_OSDMAP_REQUIRE_JEWEL)) { |
9f95a23c | 972 | new_require_osd_release = ceph_release_t::jewel; |
31f18b77 | 973 | } else { |
9f95a23c | 974 | new_require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
975 | } |
976 | } | |
81eedcae TL |
977 | if (struct_v >= 8) { |
978 | decode(new_crush_node_flags, bl); | |
979 | } | |
980 | if (struct_v >= 9) { | |
981 | decode(new_device_class_flags, bl); | |
982 | } | |
f67539c2 TL |
983 | if (struct_v >= 10) { |
984 | decode(change_stretch_mode, bl); | |
985 | decode(new_stretch_bucket_count, bl); | |
986 | decode(new_degraded_stretch_mode, bl); | |
987 | decode(new_recovering_stretch_mode, bl); | |
988 | decode(new_stretch_mode_bucket, bl); | |
989 | decode(stretch_mode_enabled, bl); | |
990 | } | |
33c7a0ef TL |
991 | if (struct_v >= 11) { |
992 | decode(new_range_blocklist, bl); | |
993 | decode(old_range_blocklist, bl); | |
994 | } | |
7c673cae FG |
995 | DECODE_FINISH(bl); // osd-only data |
996 | } | |
997 | ||
998 | if (struct_v >= 8) { | |
999 | have_crc = true; | |
1000 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 1001 | decode(inc_crc, bl); |
7c673cae | 1002 | tail_offset = bl.get_off(); |
11fdf7f2 | 1003 | decode(full_crc, bl); |
7c673cae FG |
1004 | } else { |
1005 | have_crc = false; | |
1006 | full_crc = 0; | |
1007 | inc_crc = 0; | |
1008 | } | |
1009 | ||
1010 | DECODE_FINISH(bl); // wrapper | |
1011 | ||
1012 | if (have_crc) { | |
1013 | // verify crc | |
1014 | uint32_t actual = crc_front.crc32c(-1); | |
1015 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 1016 | ceph::buffer::list tail; |
7c673cae FG |
1017 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
1018 | actual = tail.crc32c(actual); | |
1019 | } | |
1020 | if (inc_crc != actual) { | |
1021 | ostringstream ss; | |
1022 | ss << "bad crc, actual " << actual << " != expected " << inc_crc; | |
1023 | string s = ss.str(); | |
9f95a23c | 1024 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
1025 | } |
1026 | } | |
1027 | } | |
1028 | ||
1029 | void OSDMap::Incremental::dump(Formatter *f) const | |
1030 | { | |
1031 | f->dump_int("epoch", epoch); | |
1032 | f->dump_stream("fsid") << fsid; | |
1033 | f->dump_stream("modified") << modified; | |
11fdf7f2 TL |
1034 | f->dump_stream("new_last_up_change") << new_last_up_change; |
1035 | f->dump_stream("new_last_in_change") << new_last_in_change; | |
7c673cae FG |
1036 | f->dump_int("new_pool_max", new_pool_max); |
1037 | f->dump_int("new_flags", new_flags); | |
1038 | f->dump_float("new_full_ratio", new_full_ratio); | |
1039 | f->dump_float("new_nearfull_ratio", new_nearfull_ratio); | |
1040 | f->dump_float("new_backfillfull_ratio", new_backfillfull_ratio); | |
f67539c2 TL |
1041 | f->dump_int("new_require_min_compat_client", to_integer<int>(new_require_min_compat_client)); |
1042 | f->dump_int("new_require_osd_release", to_integer<int>(new_require_osd_release)); | |
7c673cae FG |
1043 | |
1044 | if (fullmap.length()) { | |
1045 | f->open_object_section("full_map"); | |
1046 | OSDMap full; | |
9f95a23c | 1047 | ceph::buffer::list fbl = fullmap; // kludge around constness. |
11fdf7f2 | 1048 | auto p = fbl.cbegin(); |
7c673cae FG |
1049 | full.decode(p); |
1050 | full.dump(f); | |
1051 | f->close_section(); | |
1052 | } | |
1053 | if (crush.length()) { | |
1054 | f->open_object_section("crush"); | |
1055 | CrushWrapper c; | |
9f95a23c | 1056 | ceph::buffer::list tbl = crush; // kludge around constness. |
11fdf7f2 | 1057 | auto p = tbl.cbegin(); |
7c673cae FG |
1058 | c.decode(p); |
1059 | c.dump(f); | |
1060 | f->close_section(); | |
1061 | } | |
1062 | ||
1063 | f->dump_int("new_max_osd", new_max_osd); | |
1064 | ||
1065 | f->open_array_section("new_pools"); | |
1066 | ||
1067 | for (const auto &new_pool : new_pools) { | |
1068 | f->open_object_section("pool"); | |
1069 | f->dump_int("pool", new_pool.first); | |
1070 | new_pool.second.dump(f); | |
1071 | f->close_section(); | |
1072 | } | |
1073 | f->close_section(); | |
1074 | f->open_array_section("new_pool_names"); | |
1075 | ||
1076 | for (const auto &new_pool_name : new_pool_names) { | |
1077 | f->open_object_section("pool_name"); | |
1078 | f->dump_int("pool", new_pool_name.first); | |
1079 | f->dump_string("name", new_pool_name.second); | |
1080 | f->close_section(); | |
1081 | } | |
1082 | f->close_section(); | |
1083 | f->open_array_section("old_pools"); | |
1084 | ||
1085 | for (const auto &old_pool : old_pools) | |
1086 | f->dump_int("pool", old_pool); | |
1087 | f->close_section(); | |
1088 | ||
1089 | f->open_array_section("new_up_osds"); | |
1090 | ||
1091 | for (const auto &upclient : new_up_client) { | |
1092 | f->open_object_section("osd"); | |
1093 | f->dump_int("osd", upclient.first); | |
11fdf7f2 TL |
1094 | f->dump_stream("public_addr") << upclient.second.legacy_addr(); |
1095 | f->dump_object("public_addrs", upclient.second); | |
1096 | if (auto p = new_up_cluster.find(upclient.first); | |
1097 | p != new_up_cluster.end()) { | |
1098 | f->dump_stream("cluster_addr") << p->second.legacy_addr(); | |
1099 | f->dump_object("cluster_addrs", p->second); | |
1100 | } | |
1101 | if (auto p = new_hb_back_up.find(upclient.first); | |
1102 | p != new_hb_back_up.end()) { | |
1103 | f->dump_object("heartbeat_back_addrs", p->second); | |
1104 | } | |
1105 | if (auto p = new_hb_front_up.find(upclient.first); | |
1106 | p != new_hb_front_up.end()) { | |
1107 | f->dump_object("heartbeat_front_addrs", p->second); | |
1108 | } | |
7c673cae FG |
1109 | f->close_section(); |
1110 | } | |
1111 | f->close_section(); | |
1112 | ||
1113 | f->open_array_section("new_weight"); | |
1114 | ||
1115 | for (const auto &weight : new_weight) { | |
1116 | f->open_object_section("osd"); | |
1117 | f->dump_int("osd", weight.first); | |
1118 | f->dump_int("weight", weight.second); | |
1119 | f->close_section(); | |
1120 | } | |
1121 | f->close_section(); | |
1122 | ||
1123 | f->open_array_section("osd_state_xor"); | |
1124 | for (const auto &ns : new_state) { | |
1125 | f->open_object_section("osd"); | |
1126 | f->dump_int("osd", ns.first); | |
1127 | set<string> st; | |
1128 | calc_state_set(new_state.find(ns.first)->second, st); | |
1129 | f->open_array_section("state_xor"); | |
1130 | for (auto &state : st) | |
1131 | f->dump_string("state", state); | |
1132 | f->close_section(); | |
c07f9fc5 | 1133 | f->close_section(); |
7c673cae FG |
1134 | } |
1135 | f->close_section(); | |
1136 | ||
1137 | f->open_array_section("new_pg_temp"); | |
1138 | ||
1139 | for (const auto &pg_temp : new_pg_temp) { | |
1140 | f->open_object_section("pg"); | |
1141 | f->dump_stream("pgid") << pg_temp.first; | |
1142 | f->open_array_section("osds"); | |
1143 | ||
1144 | for (const auto &osd : pg_temp.second) | |
1145 | f->dump_int("osd", osd); | |
1146 | f->close_section(); | |
1147 | f->close_section(); | |
1148 | } | |
1149 | f->close_section(); | |
1150 | ||
1151 | f->open_array_section("primary_temp"); | |
1152 | ||
1153 | for (const auto &primary_temp : new_primary_temp) { | |
1154 | f->dump_stream("pgid") << primary_temp.first; | |
1155 | f->dump_int("osd", primary_temp.second); | |
1156 | } | |
1157 | f->close_section(); // primary_temp | |
1158 | ||
1159 | f->open_array_section("new_pg_upmap"); | |
1160 | for (auto& i : new_pg_upmap) { | |
1161 | f->open_object_section("mapping"); | |
1162 | f->dump_stream("pgid") << i.first; | |
1163 | f->open_array_section("osds"); | |
1164 | for (auto osd : i.second) { | |
1165 | f->dump_int("osd", osd); | |
1166 | } | |
1167 | f->close_section(); | |
1168 | f->close_section(); | |
1169 | } | |
1170 | f->close_section(); | |
1171 | f->open_array_section("old_pg_upmap"); | |
1172 | for (auto& i : old_pg_upmap) { | |
1173 | f->dump_stream("pgid") << i; | |
1174 | } | |
1175 | f->close_section(); | |
1176 | ||
1177 | f->open_array_section("new_pg_upmap_items"); | |
1178 | for (auto& i : new_pg_upmap_items) { | |
1179 | f->open_object_section("mapping"); | |
1180 | f->dump_stream("pgid") << i.first; | |
1181 | f->open_array_section("mappings"); | |
1182 | for (auto& p : i.second) { | |
1183 | f->open_object_section("mapping"); | |
1184 | f->dump_int("from", p.first); | |
1185 | f->dump_int("to", p.second); | |
1186 | f->close_section(); | |
1187 | } | |
1188 | f->close_section(); | |
1189 | f->close_section(); | |
1190 | } | |
1191 | f->close_section(); | |
1192 | f->open_array_section("old_pg_upmap_items"); | |
1193 | for (auto& i : old_pg_upmap_items) { | |
1194 | f->dump_stream("pgid") << i; | |
1195 | } | |
1196 | f->close_section(); | |
1197 | ||
1198 | f->open_array_section("new_up_thru"); | |
1199 | ||
1200 | for (const auto &up_thru : new_up_thru) { | |
1201 | f->open_object_section("osd"); | |
1202 | f->dump_int("osd", up_thru.first); | |
1203 | f->dump_int("up_thru", up_thru.second); | |
1204 | f->close_section(); | |
1205 | } | |
1206 | f->close_section(); | |
1207 | ||
1208 | f->open_array_section("new_lost"); | |
1209 | ||
1210 | for (const auto &lost : new_lost) { | |
1211 | f->open_object_section("osd"); | |
1212 | f->dump_int("osd", lost.first); | |
1213 | f->dump_int("epoch_lost", lost.second); | |
1214 | f->close_section(); | |
1215 | } | |
1216 | f->close_section(); | |
1217 | ||
1218 | f->open_array_section("new_last_clean_interval"); | |
1219 | ||
1220 | for (const auto &last_clean_interval : new_last_clean_interval) { | |
1221 | f->open_object_section("osd"); | |
1222 | f->dump_int("osd", last_clean_interval.first); | |
1223 | f->dump_int("first", last_clean_interval.second.first); | |
1224 | f->dump_int("last", last_clean_interval.second.second); | |
1225 | f->close_section(); | |
1226 | } | |
1227 | f->close_section(); | |
1228 | ||
f67539c2 TL |
1229 | f->open_array_section("new_blocklist"); |
1230 | for (const auto &blist : new_blocklist) { | |
7c673cae FG |
1231 | stringstream ss; |
1232 | ss << blist.first; | |
1233 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1234 | } | |
1235 | f->close_section(); | |
f67539c2 TL |
1236 | f->open_array_section("old_blocklist"); |
1237 | for (const auto &blist : old_blocklist) | |
7c673cae FG |
1238 | f->dump_stream("addr") << blist; |
1239 | f->close_section(); | |
33c7a0ef TL |
1240 | f->open_array_section("new_range_blocklist"); |
1241 | for (const auto &blist : new_range_blocklist) { | |
1242 | stringstream ss; | |
1243 | ss << blist.first; | |
1244 | f->dump_stream(ss.str().c_str()) << blist.second; | |
1245 | } | |
1246 | f->close_section(); | |
1247 | f->open_array_section("old_range_blocklist"); | |
1248 | for (const auto &blist : old_range_blocklist) | |
1249 | f->dump_stream("addr") << blist; | |
1250 | f->close_section(); | |
7c673cae FG |
1251 | |
1252 | f->open_array_section("new_xinfo"); | |
1253 | for (const auto &xinfo : new_xinfo) { | |
1254 | f->open_object_section("xinfo"); | |
1255 | f->dump_int("osd", xinfo.first); | |
1256 | xinfo.second.dump(f); | |
1257 | f->close_section(); | |
1258 | } | |
1259 | f->close_section(); | |
1260 | ||
1261 | if (cluster_snapshot.size()) | |
1262 | f->dump_string("cluster_snapshot", cluster_snapshot); | |
1263 | ||
1264 | f->open_array_section("new_uuid"); | |
1265 | for (const auto &uuid : new_uuid) { | |
1266 | f->open_object_section("osd"); | |
1267 | f->dump_int("osd", uuid.first); | |
1268 | f->dump_stream("uuid") << uuid.second; | |
1269 | f->close_section(); | |
1270 | } | |
1271 | f->close_section(); | |
1272 | ||
1273 | OSDMap::dump_erasure_code_profiles(new_erasure_code_profiles, f); | |
1274 | f->open_array_section("old_erasure_code_profiles"); | |
1275 | for (const auto &erasure_code_profile : old_erasure_code_profiles) { | |
9f95a23c | 1276 | f->dump_string("old", erasure_code_profile); |
7c673cae FG |
1277 | } |
1278 | f->close_section(); | |
11fdf7f2 TL |
1279 | |
1280 | f->open_array_section("new_removed_snaps"); | |
1281 | for (auto& p : new_removed_snaps) { | |
1282 | f->open_object_section("pool"); | |
1283 | f->dump_int("pool", p.first); | |
1284 | f->open_array_section("snaps"); | |
1285 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1286 | f->open_object_section("interval"); | |
1287 | f->dump_unsigned("begin", q.get_start()); | |
1288 | f->dump_unsigned("length", q.get_len()); | |
1289 | f->close_section(); | |
1290 | } | |
1291 | f->close_section(); | |
1292 | f->close_section(); | |
1293 | } | |
1294 | f->close_section(); | |
1295 | f->open_array_section("new_purged_snaps"); | |
1296 | for (auto& p : new_purged_snaps) { | |
1297 | f->open_object_section("pool"); | |
1298 | f->dump_int("pool", p.first); | |
1299 | f->open_array_section("snaps"); | |
1300 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
1301 | f->open_object_section("interval"); | |
1302 | f->dump_unsigned("begin", q.get_start()); | |
1303 | f->dump_unsigned("length", q.get_len()); | |
1304 | f->close_section(); | |
1305 | } | |
1306 | f->close_section(); | |
1307 | f->close_section(); | |
1308 | } | |
81eedcae TL |
1309 | f->open_array_section("new_crush_node_flags"); |
1310 | for (auto& i : new_crush_node_flags) { | |
1311 | f->open_object_section("node"); | |
1312 | f->dump_int("id", i.first); | |
1313 | set<string> st; | |
1314 | calc_state_set(i.second, st); | |
1315 | for (auto& j : st) { | |
1316 | f->dump_string("flag", j); | |
1317 | } | |
1318 | f->close_section(); | |
1319 | } | |
1320 | f->close_section(); | |
1321 | f->open_array_section("new_device_class_flags"); | |
1322 | for (auto& i : new_device_class_flags) { | |
1323 | f->open_object_section("device_class"); | |
1324 | f->dump_int("id", i.first); | |
1325 | set<string> st; | |
1326 | calc_state_set(i.second, st); | |
1327 | for (auto& j : st) { | |
1328 | f->dump_string("flag", j); | |
1329 | } | |
1330 | f->close_section(); | |
1331 | } | |
1332 | f->close_section(); | |
f67539c2 TL |
1333 | f->open_object_section("stretch_mode"); |
1334 | { | |
1335 | f->dump_bool("change_stretch_mode", change_stretch_mode); | |
1336 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
1337 | f->dump_unsigned("new_stretch_bucket_count", new_stretch_bucket_count); | |
1338 | f->dump_unsigned("new_degraded_stretch_mode", new_degraded_stretch_mode); | |
1339 | f->dump_unsigned("new_recovering_stretch_mode", new_recovering_stretch_mode); | |
1340 | f->dump_int("new_stretch_mode_bucket", new_stretch_mode_bucket); | |
1341 | } | |
1342 | f->close_section(); | |
11fdf7f2 | 1343 | f->close_section(); |
7c673cae FG |
1344 | } |
1345 | ||
1346 | void OSDMap::Incremental::generate_test_instances(list<Incremental*>& o) | |
1347 | { | |
1348 | o.push_back(new Incremental); | |
1349 | } | |
1350 | ||
1351 | // ---------------------------------- | |
1352 | // OSDMap | |
1353 | ||
1354 | void OSDMap::set_epoch(epoch_t e) | |
1355 | { | |
1356 | epoch = e; | |
1357 | for (auto &pool : pools) | |
1358 | pool.second.last_change = e; | |
1359 | } | |
1360 | ||
33c7a0ef TL |
1361 | OSDMap::range_bits::range_bits() : ipv6(false) { |
1362 | memset(&bits, 0, sizeof(bits)); | |
1363 | } | |
1364 | ||
1365 | OSDMap::range_bits::range_bits(const entity_addr_t& addr) : ipv6(false) { | |
1366 | memset(&bits, 0, sizeof(bits)); | |
1367 | parse(addr); | |
1368 | } | |
1369 | ||
1370 | void OSDMap::range_bits::get_ipv6_bytes(unsigned const char *addr, | |
1371 | uint64_t *upper, uint64_t *lower) | |
1372 | { | |
1373 | *upper = ((uint64_t)(ntohl(*(uint32_t*)(addr)))) << 32 | | |
1374 | ((uint64_t)(ntohl(*(uint32_t*)(&addr[4])))); | |
1375 | *lower = ((uint64_t)(ntohl(*(uint32_t*)(&addr[8])))) << 32 | | |
1376 | ((uint64_t)(ntohl(*(uint32_t*)(&addr[12])))); | |
1377 | } | |
1378 | ||
1379 | void OSDMap::range_bits::parse(const entity_addr_t& addr) { | |
1380 | // parse it into meaningful data | |
1381 | if (addr.is_ipv6()) { | |
1382 | get_ipv6_bytes(addr.in6_addr().sin6_addr.s6_addr, | |
1383 | &bits.ipv6.upper_64_bits, &bits.ipv6.lower_64_bits); | |
1384 | int32_t lower_shift = std::min(128- | |
1385 | static_cast<int32_t>(addr.get_nonce()), 64); | |
1386 | int32_t upper_shift = std::max(64- //(128-b.first.get_nonce())-64 | |
1387 | static_cast<int32_t>(addr.get_nonce()), 0); | |
1388 | ||
1389 | auto get_mask = [](int32_t shift) -> uint64_t { | |
1390 | if (shift >= 0 && shift < 64) { | |
1391 | return UINT64_MAX << shift; | |
1392 | } | |
1393 | return 0; | |
1394 | }; | |
1395 | ||
1396 | bits.ipv6.lower_mask = get_mask(lower_shift); | |
1397 | bits.ipv6.upper_mask = get_mask(upper_shift); | |
1398 | ipv6 = true; | |
1399 | } else if (addr.is_ipv4()) { | |
1400 | bits.ipv4.ip_32_bits = ntohl(addr.in4_addr().sin_addr.s_addr); | |
1401 | if (addr.get_nonce() > 0) { | |
1402 | bits.ipv4.mask = UINT32_MAX << (32-addr.get_nonce()); | |
1403 | } else { | |
1404 | bits.ipv4.mask = 0; | |
1405 | } | |
1406 | } else { | |
1407 | // uh... | |
1408 | } | |
1409 | } | |
1410 | ||
1411 | bool OSDMap::range_bits::matches(const entity_addr_t& addr) const { | |
1412 | if (addr.is_ipv4() && !ipv6) { | |
1413 | return ((ntohl(addr.in4_addr().sin_addr.s_addr) & bits.ipv4.mask) == | |
1414 | (bits.ipv4.ip_32_bits & bits.ipv4.mask)); | |
1415 | } else if (addr.is_ipv6() && ipv6) { | |
1416 | uint64_t upper_64, lower_64; | |
1417 | get_ipv6_bytes(addr.in6_addr().sin6_addr.s6_addr, &upper_64, &lower_64); | |
1418 | return (((upper_64 & bits.ipv6.upper_mask) == | |
1419 | (bits.ipv6.upper_64_bits & bits.ipv6.upper_mask)) && | |
1420 | ((lower_64 & bits.ipv6.lower_mask) == | |
1421 | (bits.ipv6.lower_64_bits & bits.ipv6.lower_mask))); | |
1422 | } | |
1423 | return false; | |
1424 | } | |
1425 | ||
1426 | bool OSDMap::is_blocklisted(const entity_addr_t& orig, CephContext *cct) const | |
7c673cae | 1427 | { |
33c7a0ef TL |
1428 | if (cct) ldout(cct, 25) << "is_blocklisted: " << orig << dendl; |
1429 | if (blocklist.empty() && range_blocklist.empty()) { | |
1430 | if (cct) ldout(cct, 30) << "not blocklisted: " << orig << dendl; | |
7c673cae | 1431 | return false; |
11fdf7f2 TL |
1432 | } |
1433 | ||
f67539c2 | 1434 | // all blocklist entries are type ANY for nautilus+ |
11fdf7f2 TL |
1435 | // FIXME: avoid this copy! |
1436 | entity_addr_t a = orig; | |
9f95a23c | 1437 | if (require_osd_release < ceph_release_t::nautilus) { |
11fdf7f2 TL |
1438 | a.set_type(entity_addr_t::TYPE_LEGACY); |
1439 | } else { | |
1440 | a.set_type(entity_addr_t::TYPE_ANY); | |
1441 | } | |
7c673cae FG |
1442 | |
1443 | // this specific instance? | |
f67539c2 | 1444 | if (blocklist.count(a)) { |
33c7a0ef | 1445 | if (cct) ldout(cct, 20) << "blocklist contains " << a << dendl; |
7c673cae | 1446 | return true; |
11fdf7f2 | 1447 | } |
7c673cae | 1448 | |
f67539c2 | 1449 | // is entire ip blocklisted? |
7c673cae | 1450 | if (a.is_ip()) { |
11fdf7f2 TL |
1451 | a.set_port(0); |
1452 | a.set_nonce(0); | |
f67539c2 | 1453 | if (blocklist.count(a)) { |
33c7a0ef | 1454 | if (cct) ldout(cct, 20) << "blocklist contains " << a << dendl; |
11fdf7f2 TL |
1455 | return true; |
1456 | } | |
1457 | } | |
1458 | ||
33c7a0ef TL |
1459 | // is it in a blocklisted range? |
1460 | for (const auto& i : calculated_ranges) { | |
1461 | bool blocked = i.second.matches(a); | |
1462 | if (blocked) { | |
1463 | if (cct) ldout(cct, 20) << "range_blocklist contains " << a << dendl; | |
1464 | return true; | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | if (cct) ldout(cct, 25) << "not blocklisted: " << orig << dendl; | |
11fdf7f2 TL |
1469 | return false; |
1470 | } | |
1471 | ||
33c7a0ef | 1472 | bool OSDMap::is_blocklisted(const entity_addrvec_t& av, CephContext *cct) const |
11fdf7f2 | 1473 | { |
33c7a0ef | 1474 | if (blocklist.empty() && range_blocklist.empty()) |
11fdf7f2 TL |
1475 | return false; |
1476 | ||
1477 | for (auto& a : av.v) { | |
33c7a0ef | 1478 | if (is_blocklisted(a, cct)) { |
7c673cae FG |
1479 | return true; |
1480 | } | |
1481 | } | |
1482 | ||
1483 | return false; | |
1484 | } | |
1485 | ||
33c7a0ef TL |
1486 | void OSDMap::get_blocklist(list<pair<entity_addr_t,utime_t> > *bl, |
1487 | std::list<std::pair<entity_addr_t,utime_t> > *rl) const | |
7c673cae | 1488 | { |
f67539c2 | 1489 | std::copy(blocklist.begin(), blocklist.end(), std::back_inserter(*bl)); |
33c7a0ef TL |
1490 | std::copy(range_blocklist.begin(), range_blocklist.end(), |
1491 | std::back_inserter(*rl)); | |
7c673cae FG |
1492 | } |
1493 | ||
33c7a0ef TL |
1494 | void OSDMap::get_blocklist(std::set<entity_addr_t> *bl, |
1495 | std::set<entity_addr_t> *rl) const | |
31f18b77 | 1496 | { |
f67539c2 | 1497 | for (const auto &i : blocklist) { |
31f18b77 FG |
1498 | bl->insert(i.first); |
1499 | } | |
33c7a0ef TL |
1500 | for (const auto &i : range_blocklist) { |
1501 | rl->insert(i.first); | |
1502 | } | |
31f18b77 FG |
1503 | } |
1504 | ||
7c673cae FG |
1505 | void OSDMap::set_max_osd(int m) |
1506 | { | |
7c673cae | 1507 | max_osd = m; |
f67539c2 TL |
1508 | osd_state.resize(max_osd, 0); |
1509 | osd_weight.resize(max_osd, CEPH_OSD_OUT); | |
1510 | osd_info.resize(max_osd); | |
1511 | osd_xinfo.resize(max_osd); | |
1512 | osd_addrs->client_addrs.resize(max_osd); | |
1513 | osd_addrs->cluster_addrs.resize(max_osd); | |
1514 | osd_addrs->hb_back_addrs.resize(max_osd); | |
1515 | osd_addrs->hb_front_addrs.resize(max_osd); | |
1516 | osd_uuid->resize(max_osd); | |
7c673cae | 1517 | if (osd_primary_affinity) |
f67539c2 | 1518 | osd_primary_affinity->resize(max_osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); |
7c673cae FG |
1519 | |
1520 | calc_num_osds(); | |
1521 | } | |
1522 | ||
1523 | int OSDMap::calc_num_osds() | |
1524 | { | |
1525 | num_osd = 0; | |
1526 | num_up_osd = 0; | |
1527 | num_in_osd = 0; | |
1528 | for (int i=0; i<max_osd; i++) { | |
1529 | if (osd_state[i] & CEPH_OSD_EXISTS) { | |
1530 | ++num_osd; | |
1531 | if (osd_state[i] & CEPH_OSD_UP) { | |
1532 | ++num_up_osd; | |
1533 | } | |
1534 | if (get_weight(i) != CEPH_OSD_OUT) { | |
1535 | ++num_in_osd; | |
1536 | } | |
1537 | } | |
1538 | } | |
1539 | return num_osd; | |
1540 | } | |
1541 | ||
3efd9988 FG |
1542 | void OSDMap::get_full_pools(CephContext *cct, |
1543 | set<int64_t> *full, | |
1544 | set<int64_t> *backfillfull, | |
1545 | set<int64_t> *nearfull) const | |
7c673cae | 1546 | { |
11fdf7f2 TL |
1547 | ceph_assert(full); |
1548 | ceph_assert(backfillfull); | |
1549 | ceph_assert(nearfull); | |
3efd9988 FG |
1550 | full->clear(); |
1551 | backfillfull->clear(); | |
1552 | nearfull->clear(); | |
1553 | ||
1554 | vector<int> full_osds; | |
1555 | vector<int> backfillfull_osds; | |
1556 | vector<int> nearfull_osds; | |
7c673cae FG |
1557 | for (int i = 0; i < max_osd; ++i) { |
1558 | if (exists(i) && is_up(i) && is_in(i)) { | |
1559 | if (osd_state[i] & CEPH_OSD_FULL) | |
3efd9988 | 1560 | full_osds.push_back(i); |
7c673cae | 1561 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) |
3efd9988 | 1562 | backfillfull_osds.push_back(i); |
7c673cae | 1563 | else if (osd_state[i] & CEPH_OSD_NEARFULL) |
3efd9988 | 1564 | nearfull_osds.push_back(i); |
7c673cae FG |
1565 | } |
1566 | } | |
3efd9988 FG |
1567 | |
1568 | for (auto i: full_osds) { | |
1569 | get_pool_ids_by_osd(cct, i, full); | |
1570 | } | |
1571 | for (auto i: backfillfull_osds) { | |
1572 | get_pool_ids_by_osd(cct, i, backfillfull); | |
1573 | } | |
1574 | for (auto i: nearfull_osds) { | |
1575 | get_pool_ids_by_osd(cct, i, nearfull); | |
1576 | } | |
7c673cae FG |
1577 | } |
1578 | ||
31f18b77 FG |
1579 | void OSDMap::get_full_osd_counts(set<int> *full, set<int> *backfill, |
1580 | set<int> *nearfull) const | |
1581 | { | |
1582 | full->clear(); | |
1583 | backfill->clear(); | |
1584 | nearfull->clear(); | |
1585 | for (int i = 0; i < max_osd; ++i) { | |
1586 | if (exists(i) && is_up(i) && is_in(i)) { | |
1587 | if (osd_state[i] & CEPH_OSD_FULL) | |
1588 | full->emplace(i); | |
1589 | else if (osd_state[i] & CEPH_OSD_BACKFILLFULL) | |
1590 | backfill->emplace(i); | |
1591 | else if (osd_state[i] & CEPH_OSD_NEARFULL) | |
1592 | nearfull->emplace(i); | |
1593 | } | |
1594 | } | |
1595 | } | |
1596 | ||
7c673cae FG |
1597 | void OSDMap::get_all_osds(set<int32_t>& ls) const |
1598 | { | |
1599 | for (int i=0; i<max_osd; i++) | |
1600 | if (exists(i)) | |
1601 | ls.insert(i); | |
1602 | } | |
1603 | ||
1604 | void OSDMap::get_up_osds(set<int32_t>& ls) const | |
1605 | { | |
1606 | for (int i = 0; i < max_osd; i++) { | |
1607 | if (is_up(i)) | |
1608 | ls.insert(i); | |
1609 | } | |
1610 | } | |
1611 | ||
81eedcae | 1612 | void OSDMap::get_out_existing_osds(set<int32_t>& ls) const |
31f18b77 FG |
1613 | { |
1614 | for (int i = 0; i < max_osd; i++) { | |
81eedcae | 1615 | if (exists(i) && get_weight(i) == CEPH_OSD_OUT) |
31f18b77 FG |
1616 | ls.insert(i); |
1617 | } | |
1618 | } | |
1619 | ||
11fdf7f2 TL |
1620 | void OSDMap::get_flag_set(set<string> *flagset) const |
1621 | { | |
1622 | for (unsigned i = 0; i < sizeof(flags) * 8; ++i) { | |
1623 | if (flags & (1<<i)) { | |
1624 | flagset->insert(get_flag_string(flags & (1<<i))); | |
1625 | } | |
1626 | } | |
1627 | } | |
1628 | ||
7c673cae FG |
1629 | void OSDMap::calc_state_set(int state, set<string>& st) |
1630 | { | |
1631 | unsigned t = state; | |
1632 | for (unsigned s = 1; t; s <<= 1) { | |
1633 | if (t & s) { | |
1634 | t &= ~s; | |
1635 | st.insert(ceph_osd_state_name(s)); | |
1636 | } | |
1637 | } | |
1638 | } | |
1639 | ||
1640 | void OSDMap::adjust_osd_weights(const map<int,double>& weights, Incremental& inc) const | |
1641 | { | |
1642 | float max = 0; | |
1643 | for (const auto &weight : weights) { | |
1644 | if (weight.second > max) | |
1645 | max = weight.second; | |
1646 | } | |
1647 | ||
1648 | for (const auto &weight : weights) { | |
1649 | inc.new_weight[weight.first] = (unsigned)((weight.second / max) * CEPH_OSD_IN); | |
1650 | } | |
1651 | } | |
1652 | ||
1653 | int OSDMap::identify_osd(const entity_addr_t& addr) const | |
1654 | { | |
1655 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1656 | if (exists(i) && (get_addrs(i).contains(addr) || |
1657 | get_cluster_addrs(i).contains(addr))) | |
7c673cae FG |
1658 | return i; |
1659 | return -1; | |
1660 | } | |
1661 | ||
1662 | int OSDMap::identify_osd(const uuid_d& u) const | |
1663 | { | |
1664 | for (int i=0; i<max_osd; i++) | |
1665 | if (exists(i) && get_uuid(i) == u) | |
1666 | return i; | |
1667 | return -1; | |
1668 | } | |
1669 | ||
1670 | int OSDMap::identify_osd_on_all_channels(const entity_addr_t& addr) const | |
1671 | { | |
1672 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1673 | if (exists(i) && (get_addrs(i).contains(addr) || |
1674 | get_cluster_addrs(i).contains(addr) || | |
1675 | get_hb_back_addrs(i).contains(addr) || | |
1676 | get_hb_front_addrs(i).contains(addr))) | |
7c673cae FG |
1677 | return i; |
1678 | return -1; | |
1679 | } | |
1680 | ||
1681 | int OSDMap::find_osd_on_ip(const entity_addr_t& ip) const | |
1682 | { | |
1683 | for (int i=0; i<max_osd; i++) | |
11fdf7f2 TL |
1684 | if (exists(i) && (get_addrs(i).is_same_host(ip) || |
1685 | get_cluster_addrs(i).is_same_host(ip))) | |
7c673cae FG |
1686 | return i; |
1687 | return -1; | |
1688 | } | |
1689 | ||
1690 | ||
1691 | uint64_t OSDMap::get_features(int entity_type, uint64_t *pmask) const | |
1692 | { | |
1693 | uint64_t features = 0; // things we actually have | |
1694 | uint64_t mask = 0; // things we could have | |
1695 | ||
1696 | if (crush->has_nondefault_tunables()) | |
1697 | features |= CEPH_FEATURE_CRUSH_TUNABLES; | |
1698 | if (crush->has_nondefault_tunables2()) | |
1699 | features |= CEPH_FEATURE_CRUSH_TUNABLES2; | |
1700 | if (crush->has_nondefault_tunables3()) | |
1701 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1702 | if (crush->has_v4_buckets()) | |
1703 | features |= CEPH_FEATURE_CRUSH_V4; | |
1704 | if (crush->has_nondefault_tunables5()) | |
1705 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
c07f9fc5 FG |
1706 | if (crush->has_incompat_choose_args()) { |
1707 | features |= CEPH_FEATUREMASK_CRUSH_CHOOSE_ARGS; | |
1708 | } | |
7c673cae FG |
1709 | mask |= CEPH_FEATURES_CRUSH; |
1710 | ||
1711 | if (!pg_upmap.empty() || !pg_upmap_items.empty()) | |
1712 | features |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1713 | mask |= CEPH_FEATUREMASK_OSDMAP_PG_UPMAP; | |
1714 | ||
1715 | for (auto &pool: pools) { | |
1716 | if (pool.second.has_flag(pg_pool_t::FLAG_HASHPSPOOL)) { | |
1717 | features |= CEPH_FEATURE_OSDHASHPSPOOL; | |
1718 | } | |
7c673cae FG |
1719 | if (!pool.second.tiers.empty() || |
1720 | pool.second.is_tier()) { | |
1721 | features |= CEPH_FEATURE_OSD_CACHEPOOL; | |
1722 | } | |
20effc67 | 1723 | int ruleid = pool.second.get_crush_rule(); |
7c673cae FG |
1724 | if (ruleid >= 0) { |
1725 | if (crush->is_v2_rule(ruleid)) | |
1726 | features |= CEPH_FEATURE_CRUSH_V2; | |
1727 | if (crush->is_v3_rule(ruleid)) | |
1728 | features |= CEPH_FEATURE_CRUSH_TUNABLES3; | |
1729 | if (crush->is_v5_rule(ruleid)) | |
1730 | features |= CEPH_FEATURE_CRUSH_TUNABLES5; | |
1731 | } | |
1732 | } | |
7c673cae | 1733 | mask |= CEPH_FEATURE_OSDHASHPSPOOL | CEPH_FEATURE_OSD_CACHEPOOL; |
7c673cae FG |
1734 | |
1735 | if (osd_primary_affinity) { | |
1736 | for (int i = 0; i < max_osd; ++i) { | |
1737 | if ((*osd_primary_affinity)[i] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
1738 | features |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1739 | break; | |
1740 | } | |
1741 | } | |
1742 | } | |
1743 | mask |= CEPH_FEATURE_OSD_PRIMARY_AFFINITY; | |
1744 | ||
1745 | if (entity_type == CEPH_ENTITY_TYPE_OSD) { | |
1746 | const uint64_t jewel_features = CEPH_FEATURE_SERVER_JEWEL; | |
9f95a23c | 1747 | if (require_osd_release >= ceph_release_t::jewel) { |
7c673cae FG |
1748 | features |= jewel_features; |
1749 | } | |
1750 | mask |= jewel_features; | |
1751 | ||
1752 | const uint64_t kraken_features = CEPH_FEATUREMASK_SERVER_KRAKEN | |
1753 | | CEPH_FEATURE_MSG_ADDR2; | |
9f95a23c | 1754 | if (require_osd_release >= ceph_release_t::kraken) { |
7c673cae FG |
1755 | features |= kraken_features; |
1756 | } | |
1757 | mask |= kraken_features; | |
f67539c2 TL |
1758 | |
1759 | if (stretch_mode_enabled) { | |
1760 | features |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1761 | mask |= CEPH_FEATUREMASK_STRETCH_MODE; | |
1762 | } | |
7c673cae FG |
1763 | } |
1764 | ||
9f95a23c | 1765 | if (require_min_compat_client >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
1766 | // if min_compat_client is >= nautilus, require v2 cephx signatures |
1767 | // from everyone | |
1768 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
9f95a23c | 1769 | } else if (require_osd_release >= ceph_release_t::nautilus && |
11fdf7f2 TL |
1770 | entity_type == CEPH_ENTITY_TYPE_OSD) { |
1771 | // if osds are >= nautilus, at least require the signatures from them | |
1772 | features |= CEPH_FEATUREMASK_CEPHX_V2; | |
1773 | } | |
1774 | mask |= CEPH_FEATUREMASK_CEPHX_V2; | |
1775 | ||
7c673cae FG |
1776 | if (pmask) |
1777 | *pmask = mask; | |
1778 | return features; | |
1779 | } | |
1780 | ||
9f95a23c | 1781 | ceph_release_t OSDMap::get_min_compat_client() const |
7c673cae FG |
1782 | { |
1783 | uint64_t f = get_features(CEPH_ENTITY_TYPE_CLIENT, nullptr); | |
1784 | ||
1785 | if (HAVE_FEATURE(f, OSDMAP_PG_UPMAP) || // v12.0.0-1733-g27d6f43 | |
31f18b77 | 1786 | HAVE_FEATURE(f, CRUSH_CHOOSE_ARGS)) { // v12.0.1-2172-gef1ef28 |
9f95a23c | 1787 | return ceph_release_t::luminous; // v12.2.0 |
7c673cae FG |
1788 | } |
1789 | if (HAVE_FEATURE(f, CRUSH_TUNABLES5)) { // v10.0.0-612-g043a737 | |
9f95a23c | 1790 | return ceph_release_t::jewel; // v10.2.0 |
7c673cae FG |
1791 | } |
1792 | if (HAVE_FEATURE(f, CRUSH_V4)) { // v0.91-678-g325fc56 | |
9f95a23c | 1793 | return ceph_release_t::hammer; // v0.94.0 |
7c673cae FG |
1794 | } |
1795 | if (HAVE_FEATURE(f, OSD_PRIMARY_AFFINITY) || // v0.76-553-gf825624 | |
1796 | HAVE_FEATURE(f, CRUSH_TUNABLES3) || // v0.76-395-ge20a55d | |
7c673cae | 1797 | HAVE_FEATURE(f, OSD_CACHEPOOL)) { // v0.67-401-gb91c1c5 |
9f95a23c | 1798 | return ceph_release_t::firefly; // v0.80.0 |
7c673cae FG |
1799 | } |
1800 | if (HAVE_FEATURE(f, CRUSH_TUNABLES2) || // v0.54-684-g0cc47ff | |
1801 | HAVE_FEATURE(f, OSDHASHPSPOOL)) { // v0.57-398-g8cc2b0f | |
9f95a23c | 1802 | return ceph_release_t::dumpling; // v0.67.0 |
7c673cae FG |
1803 | } |
1804 | if (HAVE_FEATURE(f, CRUSH_TUNABLES)) { // v0.48argonaut-206-g6f381af | |
9f95a23c | 1805 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae | 1806 | } |
9f95a23c | 1807 | return ceph_release_t::argonaut; // v0.48argonaut-206-g6f381af |
7c673cae FG |
1808 | } |
1809 | ||
9f95a23c | 1810 | ceph_release_t OSDMap::get_require_min_compat_client() const |
11fdf7f2 TL |
1811 | { |
1812 | return require_min_compat_client; | |
1813 | } | |
1814 | ||
7c673cae FG |
1815 | void OSDMap::_calc_up_osd_features() |
1816 | { | |
1817 | bool first = true; | |
1818 | cached_up_osd_features = 0; | |
1819 | for (int osd = 0; osd < max_osd; ++osd) { | |
1820 | if (!is_up(osd)) | |
1821 | continue; | |
1822 | const osd_xinfo_t &xi = get_xinfo(osd); | |
3efd9988 FG |
1823 | if (xi.features == 0) |
1824 | continue; // bogus xinfo, maybe #20751 or similar, skipping | |
7c673cae FG |
1825 | if (first) { |
1826 | cached_up_osd_features = xi.features; | |
1827 | first = false; | |
1828 | } else { | |
1829 | cached_up_osd_features &= xi.features; | |
1830 | } | |
1831 | } | |
1832 | } | |
1833 | ||
1834 | uint64_t OSDMap::get_up_osd_features() const | |
1835 | { | |
1836 | return cached_up_osd_features; | |
1837 | } | |
1838 | ||
1839 | void OSDMap::dedup(const OSDMap *o, OSDMap *n) | |
1840 | { | |
11fdf7f2 | 1841 | using ceph::encode; |
7c673cae FG |
1842 | if (o->epoch == n->epoch) |
1843 | return; | |
1844 | ||
1845 | int diff = 0; | |
1846 | ||
1847 | // do addrs match? | |
1848 | if (o->max_osd != n->max_osd) | |
1849 | diff++; | |
1850 | for (int i = 0; i < o->max_osd && i < n->max_osd; i++) { | |
11fdf7f2 TL |
1851 | if ( n->osd_addrs->client_addrs[i] && o->osd_addrs->client_addrs[i] && |
1852 | *n->osd_addrs->client_addrs[i] == *o->osd_addrs->client_addrs[i]) | |
1853 | n->osd_addrs->client_addrs[i] = o->osd_addrs->client_addrs[i]; | |
7c673cae FG |
1854 | else |
1855 | diff++; | |
11fdf7f2 TL |
1856 | if ( n->osd_addrs->cluster_addrs[i] && o->osd_addrs->cluster_addrs[i] && |
1857 | *n->osd_addrs->cluster_addrs[i] == *o->osd_addrs->cluster_addrs[i]) | |
1858 | n->osd_addrs->cluster_addrs[i] = o->osd_addrs->cluster_addrs[i]; | |
7c673cae FG |
1859 | else |
1860 | diff++; | |
11fdf7f2 TL |
1861 | if ( n->osd_addrs->hb_back_addrs[i] && o->osd_addrs->hb_back_addrs[i] && |
1862 | *n->osd_addrs->hb_back_addrs[i] == *o->osd_addrs->hb_back_addrs[i]) | |
1863 | n->osd_addrs->hb_back_addrs[i] = o->osd_addrs->hb_back_addrs[i]; | |
7c673cae FG |
1864 | else |
1865 | diff++; | |
11fdf7f2 TL |
1866 | if ( n->osd_addrs->hb_front_addrs[i] && o->osd_addrs->hb_front_addrs[i] && |
1867 | *n->osd_addrs->hb_front_addrs[i] == *o->osd_addrs->hb_front_addrs[i]) | |
1868 | n->osd_addrs->hb_front_addrs[i] = o->osd_addrs->hb_front_addrs[i]; | |
7c673cae FG |
1869 | else |
1870 | diff++; | |
1871 | } | |
1872 | if (diff == 0) { | |
1873 | // zoinks, no differences at all! | |
1874 | n->osd_addrs = o->osd_addrs; | |
1875 | } | |
1876 | ||
1877 | // does crush match? | |
9f95a23c | 1878 | ceph::buffer::list oc, nc; |
11fdf7f2 TL |
1879 | encode(*o->crush, oc, CEPH_FEATURES_SUPPORTED_DEFAULT); |
1880 | encode(*n->crush, nc, CEPH_FEATURES_SUPPORTED_DEFAULT); | |
7c673cae FG |
1881 | if (oc.contents_equal(nc)) { |
1882 | n->crush = o->crush; | |
1883 | } | |
1884 | ||
1885 | // does pg_temp match? | |
31f18b77 FG |
1886 | if (*o->pg_temp == *n->pg_temp) |
1887 | n->pg_temp = o->pg_temp; | |
7c673cae FG |
1888 | |
1889 | // does primary_temp match? | |
1890 | if (o->primary_temp->size() == n->primary_temp->size()) { | |
1891 | if (*o->primary_temp == *n->primary_temp) | |
1892 | n->primary_temp = o->primary_temp; | |
1893 | } | |
1894 | ||
1895 | // do uuids match? | |
1896 | if (o->osd_uuid->size() == n->osd_uuid->size() && | |
1897 | *o->osd_uuid == *n->osd_uuid) | |
1898 | n->osd_uuid = o->osd_uuid; | |
1899 | } | |
1900 | ||
1901 | void OSDMap::clean_temps(CephContext *cct, | |
11fdf7f2 TL |
1902 | const OSDMap& oldmap, |
1903 | const OSDMap& nextmap, | |
1904 | Incremental *pending_inc) | |
7c673cae FG |
1905 | { |
1906 | ldout(cct, 10) << __func__ << dendl; | |
7c673cae | 1907 | |
11fdf7f2 | 1908 | for (auto pg : *nextmap.pg_temp) { |
7c673cae FG |
1909 | // if pool does not exist, remove any existing pg_temps associated with |
1910 | // it. we don't care about pg_temps on the pending_inc either; if there | |
1911 | // are new_pg_temp entries on the pending, clear them out just as well. | |
11fdf7f2 | 1912 | if (!nextmap.have_pg_pool(pg.first.pool())) { |
7c673cae FG |
1913 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first |
1914 | << " for nonexistent pool " << pg.first.pool() << dendl; | |
1915 | pending_inc->new_pg_temp[pg.first].clear(); | |
1916 | continue; | |
1917 | } | |
20effc67 TL |
1918 | if (!nextmap.pg_exists(pg.first)) { |
1919 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1920 | << " for nonexistent pg " << dendl; | |
1921 | pending_inc->new_pg_temp[pg.first].clear(); | |
1922 | continue; | |
1923 | } | |
7c673cae FG |
1924 | // all osds down? |
1925 | unsigned num_up = 0; | |
1926 | for (auto o : pg.second) { | |
11fdf7f2 | 1927 | if (!nextmap.is_down(o)) { |
7c673cae FG |
1928 | ++num_up; |
1929 | break; | |
1930 | } | |
1931 | } | |
1932 | if (num_up == 0) { | |
1933 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first | |
1934 | << " with all down osds" << pg.second << dendl; | |
1935 | pending_inc->new_pg_temp[pg.first].clear(); | |
1936 | continue; | |
1937 | } | |
1938 | // redundant pg_temp? | |
1939 | vector<int> raw_up; | |
1940 | int primary; | |
11fdf7f2 | 1941 | nextmap.pg_to_raw_up(pg.first, &raw_up, &primary); |
91327a77 | 1942 | bool remove = false; |
11fdf7f2 | 1943 | if (raw_up == pg.second) { |
7c673cae FG |
1944 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1945 | << pg.second << " that matches raw_up mapping" << dendl; | |
91327a77 AA |
1946 | remove = true; |
1947 | } | |
1948 | // oversized pg_temp? | |
11fdf7f2 | 1949 | if (pg.second.size() > nextmap.get_pg_pool(pg.first.pool())->get_size()) { |
91327a77 AA |
1950 | ldout(cct, 10) << __func__ << " removing pg_temp " << pg.first << " " |
1951 | << pg.second << " exceeds pool size" << dendl; | |
1952 | remove = true; | |
1953 | } | |
1954 | if (remove) { | |
11fdf7f2 | 1955 | if (oldmap.pg_temp->count(pg.first)) |
7c673cae FG |
1956 | pending_inc->new_pg_temp[pg.first].clear(); |
1957 | else | |
1958 | pending_inc->new_pg_temp.erase(pg.first); | |
1959 | } | |
1960 | } | |
1961 | ||
11fdf7f2 | 1962 | for (auto &pg : *nextmap.primary_temp) { |
7c673cae | 1963 | // primary down? |
11fdf7f2 | 1964 | if (nextmap.is_down(pg.second)) { |
7c673cae FG |
1965 | ldout(cct, 10) << __func__ << " removing primary_temp " << pg.first |
1966 | << " to down " << pg.second << dendl; | |
1967 | pending_inc->new_primary_temp[pg.first] = -1; | |
1968 | continue; | |
1969 | } | |
1970 | // redundant primary_temp? | |
1971 | vector<int> real_up, templess_up; | |
1972 | int real_primary, templess_primary; | |
1973 | pg_t pgid = pg.first; | |
11fdf7f2 TL |
1974 | nextmap.pg_to_acting_osds(pgid, &real_up, &real_primary); |
1975 | nextmap.pg_to_raw_up(pgid, &templess_up, &templess_primary); | |
7c673cae FG |
1976 | if (real_primary == templess_primary){ |
1977 | ldout(cct, 10) << __func__ << " removing primary_temp " | |
1978 | << pgid << " -> " << real_primary | |
1979 | << " (unnecessary/redundant)" << dendl; | |
11fdf7f2 | 1980 | if (oldmap.primary_temp->count(pgid)) |
7c673cae FG |
1981 | pending_inc->new_primary_temp[pgid] = -1; |
1982 | else | |
1983 | pending_inc->new_primary_temp.erase(pgid); | |
1984 | } | |
1985 | } | |
1986 | } | |
1987 | ||
494da23a | 1988 | void OSDMap::get_upmap_pgs(vector<pg_t> *upmap_pgs) const |
94b18763 | 1989 | { |
494da23a TL |
1990 | upmap_pgs->reserve(pg_upmap.size() + pg_upmap_items.size()); |
1991 | for (auto& p : pg_upmap) | |
1992 | upmap_pgs->push_back(p.first); | |
1993 | for (auto& p : pg_upmap_items) | |
1994 | upmap_pgs->push_back(p.first); | |
1995 | } | |
94b18763 | 1996 | |
494da23a TL |
1997 | bool OSDMap::check_pg_upmaps( |
1998 | CephContext *cct, | |
1999 | const vector<pg_t>& to_check, | |
2000 | vector<pg_t> *to_cancel, | |
2001 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> *to_remap) const | |
2002 | { | |
2003 | bool any_change = false; | |
2004 | map<int, map<int, float>> rule_weight_map; | |
28e407b8 | 2005 | for (auto& pg : to_check) { |
494da23a | 2006 | const pg_pool_t *pi = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2007 | if (!pi || pg.ps() >= pi->get_pg_num_pending()) { |
2008 | ldout(cct, 0) << __func__ << " pg " << pg << " is gone or merge source" | |
2009 | << dendl; | |
494da23a | 2010 | to_cancel->push_back(pg); |
11fdf7f2 TL |
2011 | continue; |
2012 | } | |
2013 | if (pi->is_pending_merge(pg, nullptr)) { | |
2014 | ldout(cct, 0) << __func__ << " pg " << pg << " is pending merge" | |
2015 | << dendl; | |
494da23a | 2016 | to_cancel->push_back(pg); |
94b18763 FG |
2017 | continue; |
2018 | } | |
494da23a TL |
2019 | vector<int> raw, up; |
2020 | pg_to_raw_upmap(pg, &raw, &up); | |
494da23a TL |
2021 | auto crush_rule = get_pg_pool_crush_rule(pg); |
2022 | auto r = crush->verify_upmap(cct, | |
2023 | crush_rule, | |
2024 | get_pg_pool_size(pg), | |
2025 | up); | |
a8e16298 TL |
2026 | if (r < 0) { |
2027 | ldout(cct, 0) << __func__ << " verify_upmap of pg " << pg | |
2028 | << " returning " << r | |
2029 | << dendl; | |
494da23a | 2030 | to_cancel->push_back(pg); |
a8e16298 TL |
2031 | continue; |
2032 | } | |
2033 | // below we check against crush-topology changing.. | |
28e407b8 AA |
2034 | map<int, float> weight_map; |
2035 | auto it = rule_weight_map.find(crush_rule); | |
2036 | if (it == rule_weight_map.end()) { | |
494da23a | 2037 | auto r = crush->get_rule_weight_osd_map(crush_rule, &weight_map); |
28e407b8 AA |
2038 | if (r < 0) { |
2039 | lderr(cct) << __func__ << " unable to get crush weight_map for " | |
494da23a TL |
2040 | << "crush_rule " << crush_rule |
2041 | << dendl; | |
28e407b8 AA |
2042 | continue; |
2043 | } | |
2044 | rule_weight_map[crush_rule] = weight_map; | |
2045 | } else { | |
2046 | weight_map = it->second; | |
2047 | } | |
28e407b8 | 2048 | ldout(cct, 10) << __func__ << " pg " << pg |
28e407b8 | 2049 | << " weight_map " << weight_map |
94b18763 | 2050 | << dendl; |
a8e16298 | 2051 | for (auto osd : up) { |
28e407b8 AA |
2052 | auto it = weight_map.find(osd); |
2053 | if (it == weight_map.end()) { | |
92f5a8d4 TL |
2054 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd << " is gone or has " |
2055 | << "been moved out of the specific crush-tree" | |
2056 | << dendl; | |
494da23a | 2057 | to_cancel->push_back(pg); |
94b18763 FG |
2058 | break; |
2059 | } | |
494da23a | 2060 | auto adjusted_weight = get_weightf(it->first) * it->second; |
28e407b8 | 2061 | if (adjusted_weight == 0) { |
92f5a8d4 TL |
2062 | ldout(cct, 10) << __func__ << " pg " << pg << ": osd " << osd |
2063 | << " is out/crush-out" | |
2064 | << dendl; | |
494da23a | 2065 | to_cancel->push_back(pg); |
94b18763 FG |
2066 | break; |
2067 | } | |
2068 | } | |
eafe8130 TL |
2069 | if (!to_cancel->empty() && to_cancel->back() == pg) |
2070 | continue; | |
2071 | // okay, upmap is valid | |
2072 | // continue to check if it is still necessary | |
2073 | auto i = pg_upmap.find(pg); | |
a4b75251 TL |
2074 | if (i != pg_upmap.end()) { |
2075 | if (i->second == raw) { | |
2076 | ldout(cct, 10) << "removing redundant pg_upmap " << i->first << " " | |
2077 | << i->second << dendl; | |
2078 | to_cancel->push_back(pg); | |
2079 | continue; | |
2080 | } | |
2081 | if ((int)i->second.size() != get_pg_pool_size(pg)) { | |
2082 | ldout(cct, 10) << "removing pg_upmap " << i->first << " " | |
2083 | << i->second << " != pool size " << get_pg_pool_size(pg) | |
2084 | << dendl; | |
2085 | to_cancel->push_back(pg); | |
2086 | continue; | |
2087 | } | |
eafe8130 TL |
2088 | } |
2089 | auto j = pg_upmap_items.find(pg); | |
2090 | if (j != pg_upmap_items.end()) { | |
2091 | mempool::osdmap::vector<pair<int,int>> newmap; | |
2092 | for (auto& p : j->second) { | |
2093 | if (std::find(raw.begin(), raw.end(), p.first) == raw.end()) { | |
2094 | // cancel mapping if source osd does not exist anymore | |
2095 | continue; | |
2096 | } | |
2097 | if (p.second != CRUSH_ITEM_NONE && p.second < max_osd && | |
2098 | p.second >= 0 && osd_weight[p.second] == 0) { | |
2099 | // cancel mapping if target osd is out | |
2100 | continue; | |
2101 | } | |
2102 | newmap.push_back(p); | |
2103 | } | |
2104 | if (newmap.empty()) { | |
2105 | ldout(cct, 10) << " removing no-op pg_upmap_items " | |
2106 | << j->first << " " << j->second | |
2107 | << dendl; | |
2108 | to_cancel->push_back(pg); | |
2109 | } else if (newmap != j->second) { | |
2110 | ldout(cct, 10) << " simplifying partially no-op pg_upmap_items " | |
2111 | << j->first << " " << j->second | |
2112 | << " -> " << newmap | |
2113 | << dendl; | |
2114 | to_remap->insert({pg, newmap}); | |
2115 | any_change = true; | |
2116 | } | |
2117 | } | |
28e407b8 | 2118 | } |
494da23a TL |
2119 | any_change = any_change || !to_cancel->empty(); |
2120 | return any_change; | |
2121 | } | |
2122 | ||
2123 | void OSDMap::clean_pg_upmaps( | |
2124 | CephContext *cct, | |
2125 | Incremental *pending_inc, | |
2126 | const vector<pg_t>& to_cancel, | |
2127 | const map<pg_t, mempool::osdmap::vector<pair<int,int>>>& to_remap) const | |
2128 | { | |
28e407b8 | 2129 | for (auto &pg: to_cancel) { |
494da23a TL |
2130 | auto i = pending_inc->new_pg_upmap.find(pg); |
2131 | if (i != pending_inc->new_pg_upmap.end()) { | |
2132 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2133 | << "pg_upmap entry " | |
2134 | << i->first << "->" << i->second | |
2135 | << dendl; | |
2136 | pending_inc->new_pg_upmap.erase(i); | |
94b18763 | 2137 | } |
494da23a TL |
2138 | auto j = pg_upmap.find(pg); |
2139 | if (j != pg_upmap.end()) { | |
2140 | ldout(cct, 10) << __func__ << " cancel invalid pg_upmap entry " | |
2141 | << j->first << "->" << j->second | |
2142 | << dendl; | |
2143 | pending_inc->old_pg_upmap.insert(pg); | |
2144 | } | |
2145 | auto p = pending_inc->new_pg_upmap_items.find(pg); | |
2146 | if (p != pending_inc->new_pg_upmap_items.end()) { | |
2147 | ldout(cct, 10) << __func__ << " cancel invalid pending " | |
2148 | << "pg_upmap_items entry " | |
2149 | << p->first << "->" << p->second | |
2150 | << dendl; | |
2151 | pending_inc->new_pg_upmap_items.erase(p); | |
2152 | } | |
2153 | auto q = pg_upmap_items.find(pg); | |
2154 | if (q != pg_upmap_items.end()) { | |
2155 | ldout(cct, 10) << __func__ << " cancel invalid " | |
2156 | << "pg_upmap_items entry " | |
2157 | << q->first << "->" << q->second | |
2158 | << dendl; | |
2159 | pending_inc->old_pg_upmap_items.insert(pg); | |
94b18763 FG |
2160 | } |
2161 | } | |
494da23a TL |
2162 | for (auto& i : to_remap) |
2163 | pending_inc->new_pg_upmap_items[i.first] = i.second; | |
2164 | } | |
2165 | ||
2166 | bool OSDMap::clean_pg_upmaps( | |
2167 | CephContext *cct, | |
2168 | Incremental *pending_inc) const | |
2169 | { | |
2170 | ldout(cct, 10) << __func__ << dendl; | |
2171 | vector<pg_t> to_check; | |
2172 | vector<pg_t> to_cancel; | |
2173 | map<pg_t, mempool::osdmap::vector<pair<int,int>>> to_remap; | |
2174 | ||
2175 | get_upmap_pgs(&to_check); | |
2176 | auto any_change = check_pg_upmaps(cct, to_check, &to_cancel, &to_remap); | |
2177 | clean_pg_upmaps(cct, pending_inc, to_cancel, to_remap); | |
2178 | return any_change; | |
94b18763 FG |
2179 | } |
2180 | ||
7c673cae FG |
2181 | int OSDMap::apply_incremental(const Incremental &inc) |
2182 | { | |
f67539c2 | 2183 | new_blocklist_entries = false; |
7c673cae FG |
2184 | if (inc.epoch == 1) |
2185 | fsid = inc.fsid; | |
2186 | else if (inc.fsid != fsid) | |
2187 | return -EINVAL; | |
2188 | ||
11fdf7f2 | 2189 | ceph_assert(inc.epoch == epoch+1); |
7c673cae FG |
2190 | |
2191 | epoch++; | |
2192 | modified = inc.modified; | |
2193 | ||
2194 | // full map? | |
2195 | if (inc.fullmap.length()) { | |
9f95a23c | 2196 | ceph::buffer::list bl(inc.fullmap); |
7c673cae FG |
2197 | decode(bl); |
2198 | return 0; | |
2199 | } | |
2200 | ||
2201 | // nope, incremental. | |
31f18b77 | 2202 | if (inc.new_flags >= 0) { |
7c673cae | 2203 | flags = inc.new_flags; |
31f18b77 FG |
2204 | // the below is just to cover a newly-upgraded luminous mon |
2205 | // cluster that has to set require_jewel_osds or | |
2206 | // require_kraken_osds before the osds can be upgraded to | |
2207 | // luminous. | |
2208 | if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { | |
9f95a23c TL |
2209 | if (require_osd_release < ceph_release_t::kraken) { |
2210 | require_osd_release = ceph_release_t::kraken; | |
31f18b77 FG |
2211 | } |
2212 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { | |
9f95a23c TL |
2213 | if (require_osd_release < ceph_release_t::jewel) { |
2214 | require_osd_release = ceph_release_t::jewel; | |
31f18b77 FG |
2215 | } |
2216 | } | |
2217 | } | |
7c673cae FG |
2218 | |
2219 | if (inc.new_max_osd >= 0) | |
2220 | set_max_osd(inc.new_max_osd); | |
2221 | ||
2222 | if (inc.new_pool_max != -1) | |
2223 | pool_max = inc.new_pool_max; | |
2224 | ||
2225 | for (const auto &pool : inc.new_pools) { | |
2226 | pools[pool.first] = pool.second; | |
2227 | pools[pool.first].last_change = epoch; | |
2228 | } | |
2229 | ||
11fdf7f2 TL |
2230 | new_removed_snaps = inc.new_removed_snaps; |
2231 | new_purged_snaps = inc.new_purged_snaps; | |
2232 | for (auto p = new_removed_snaps.begin(); | |
2233 | p != new_removed_snaps.end(); | |
2234 | ++p) { | |
2235 | removed_snaps_queue[p->first].union_of(p->second); | |
2236 | } | |
2237 | for (auto p = new_purged_snaps.begin(); | |
2238 | p != new_purged_snaps.end(); | |
2239 | ++p) { | |
2240 | auto q = removed_snaps_queue.find(p->first); | |
2241 | ceph_assert(q != removed_snaps_queue.end()); | |
2242 | q->second.subtract(p->second); | |
2243 | if (q->second.empty()) { | |
2244 | removed_snaps_queue.erase(q); | |
2245 | } | |
2246 | } | |
2247 | ||
2248 | if (inc.new_last_up_change != utime_t()) { | |
2249 | last_up_change = inc.new_last_up_change; | |
2250 | } | |
2251 | if (inc.new_last_in_change != utime_t()) { | |
2252 | last_in_change = inc.new_last_in_change; | |
2253 | } | |
2254 | ||
7c673cae FG |
2255 | for (const auto &pname : inc.new_pool_names) { |
2256 | auto pool_name_entry = pool_name.find(pname.first); | |
2257 | if (pool_name_entry != pool_name.end()) { | |
2258 | name_pool.erase(pool_name_entry->second); | |
2259 | pool_name_entry->second = pname.second; | |
2260 | } else { | |
2261 | pool_name[pname.first] = pname.second; | |
2262 | } | |
2263 | name_pool[pname.second] = pname.first; | |
2264 | } | |
2265 | ||
2266 | for (const auto &pool : inc.old_pools) { | |
2267 | pools.erase(pool); | |
2268 | name_pool.erase(pool_name[pool]); | |
2269 | pool_name.erase(pool); | |
2270 | } | |
2271 | ||
2272 | for (const auto &weight : inc.new_weight) { | |
2273 | set_weight(weight.first, weight.second); | |
2274 | ||
2275 | // if we are marking in, clear the AUTOOUT and NEW bits, and clear | |
2276 | // xinfo old_weight. | |
2277 | if (weight.second) { | |
2278 | osd_state[weight.first] &= ~(CEPH_OSD_AUTOOUT | CEPH_OSD_NEW); | |
2279 | osd_xinfo[weight.first].old_weight = 0; | |
2280 | } | |
2281 | } | |
2282 | ||
2283 | for (const auto &primary_affinity : inc.new_primary_affinity) { | |
2284 | set_primary_affinity(primary_affinity.first, primary_affinity.second); | |
2285 | } | |
2286 | ||
2287 | // erasure_code_profiles | |
2288 | for (const auto &profile : inc.old_erasure_code_profiles) | |
2289 | erasure_code_profiles.erase(profile); | |
2290 | ||
2291 | for (const auto &profile : inc.new_erasure_code_profiles) { | |
2292 | set_erasure_code_profile(profile.first, profile.second); | |
2293 | } | |
2294 | ||
2295 | // up/down | |
2296 | for (const auto &state : inc.new_state) { | |
2297 | const auto osd = state.first; | |
2298 | int s = state.second ? state.second : CEPH_OSD_UP; | |
2299 | if ((osd_state[osd] & CEPH_OSD_UP) && | |
2300 | (s & CEPH_OSD_UP)) { | |
2301 | osd_info[osd].down_at = epoch; | |
2302 | osd_xinfo[osd].down_stamp = modified; | |
2303 | } | |
2304 | if ((osd_state[osd] & CEPH_OSD_EXISTS) && | |
2305 | (s & CEPH_OSD_EXISTS)) { | |
2306 | // osd is destroyed; clear out anything interesting. | |
2307 | (*osd_uuid)[osd] = uuid_d(); | |
2308 | osd_info[osd] = osd_info_t(); | |
2309 | osd_xinfo[osd] = osd_xinfo_t(); | |
2310 | set_primary_affinity(osd, CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); | |
11fdf7f2 TL |
2311 | osd_addrs->client_addrs[osd].reset(new entity_addrvec_t()); |
2312 | osd_addrs->cluster_addrs[osd].reset(new entity_addrvec_t()); | |
2313 | osd_addrs->hb_front_addrs[osd].reset(new entity_addrvec_t()); | |
2314 | osd_addrs->hb_back_addrs[osd].reset(new entity_addrvec_t()); | |
7c673cae FG |
2315 | osd_state[osd] = 0; |
2316 | } else { | |
2317 | osd_state[osd] ^= s; | |
2318 | } | |
2319 | } | |
2320 | ||
2321 | for (const auto &client : inc.new_up_client) { | |
2322 | osd_state[client.first] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; | |
9f95a23c | 2323 | osd_state[client.first] &= ~CEPH_OSD_STOP; // if any |
11fdf7f2 TL |
2324 | osd_addrs->client_addrs[client.first].reset( |
2325 | new entity_addrvec_t(client.second)); | |
2326 | osd_addrs->hb_back_addrs[client.first].reset( | |
2327 | new entity_addrvec_t(inc.new_hb_back_up.find(client.first)->second)); | |
2328 | osd_addrs->hb_front_addrs[client.first].reset( | |
2329 | new entity_addrvec_t(inc.new_hb_front_up.find(client.first)->second)); | |
7c673cae FG |
2330 | |
2331 | osd_info[client.first].up_from = epoch; | |
2332 | } | |
2333 | ||
2334 | for (const auto &cluster : inc.new_up_cluster) | |
11fdf7f2 TL |
2335 | osd_addrs->cluster_addrs[cluster.first].reset( |
2336 | new entity_addrvec_t(cluster.second)); | |
7c673cae FG |
2337 | |
2338 | // info | |
2339 | for (const auto &thru : inc.new_up_thru) | |
2340 | osd_info[thru.first].up_thru = thru.second; | |
2341 | ||
2342 | for (const auto &interval : inc.new_last_clean_interval) { | |
2343 | osd_info[interval.first].last_clean_begin = interval.second.first; | |
2344 | osd_info[interval.first].last_clean_end = interval.second.second; | |
2345 | } | |
2346 | ||
2347 | for (const auto &lost : inc.new_lost) | |
2348 | osd_info[lost.first].lost_at = lost.second; | |
2349 | ||
2350 | // xinfo | |
2351 | for (const auto &xinfo : inc.new_xinfo) | |
2352 | osd_xinfo[xinfo.first] = xinfo.second; | |
2353 | ||
2354 | // uuid | |
2355 | for (const auto &uuid : inc.new_uuid) | |
2356 | (*osd_uuid)[uuid.first] = uuid.second; | |
2357 | ||
2358 | // pg rebuild | |
2359 | for (const auto &pg : inc.new_pg_temp) { | |
2360 | if (pg.second.empty()) | |
2361 | pg_temp->erase(pg.first); | |
2362 | else | |
31f18b77 FG |
2363 | pg_temp->set(pg.first, pg.second); |
2364 | } | |
2365 | if (!inc.new_pg_temp.empty()) { | |
2366 | // make sure pg_temp is efficiently stored | |
2367 | pg_temp->rebuild(); | |
7c673cae FG |
2368 | } |
2369 | ||
2370 | for (const auto &pg : inc.new_primary_temp) { | |
2371 | if (pg.second == -1) | |
2372 | primary_temp->erase(pg.first); | |
2373 | else | |
2374 | (*primary_temp)[pg.first] = pg.second; | |
2375 | } | |
2376 | ||
2377 | for (auto& p : inc.new_pg_upmap) { | |
2378 | pg_upmap[p.first] = p.second; | |
2379 | } | |
2380 | for (auto& pg : inc.old_pg_upmap) { | |
2381 | pg_upmap.erase(pg); | |
2382 | } | |
2383 | for (auto& p : inc.new_pg_upmap_items) { | |
2384 | pg_upmap_items[p.first] = p.second; | |
2385 | } | |
2386 | for (auto& pg : inc.old_pg_upmap_items) { | |
2387 | pg_upmap_items.erase(pg); | |
2388 | } | |
2389 | ||
f67539c2 TL |
2390 | // blocklist |
2391 | if (!inc.new_blocklist.empty()) { | |
2392 | blocklist.insert(inc.new_blocklist.begin(),inc.new_blocklist.end()); | |
2393 | new_blocklist_entries = true; | |
7c673cae | 2394 | } |
f67539c2 TL |
2395 | for (const auto &addr : inc.old_blocklist) |
2396 | blocklist.erase(addr); | |
7c673cae | 2397 | |
33c7a0ef TL |
2398 | for (const auto& addr_p : inc.new_range_blocklist) { |
2399 | range_blocklist.insert(addr_p); | |
2400 | calculated_ranges.emplace(addr_p.first, addr_p.first); | |
2401 | new_blocklist_entries = true; | |
2402 | } | |
2403 | for (const auto &addr : inc.old_range_blocklist) { | |
2404 | calculated_ranges.erase(addr); | |
2405 | range_blocklist.erase(addr); | |
2406 | } | |
2407 | ||
81eedcae TL |
2408 | for (auto& i : inc.new_crush_node_flags) { |
2409 | if (i.second) { | |
2410 | crush_node_flags[i.first] = i.second; | |
2411 | } else { | |
2412 | crush_node_flags.erase(i.first); | |
2413 | } | |
2414 | } | |
2415 | ||
2416 | for (auto& i : inc.new_device_class_flags) { | |
2417 | if (i.second) { | |
2418 | device_class_flags[i.first] = i.second; | |
2419 | } else { | |
2420 | device_class_flags.erase(i.first); | |
2421 | } | |
2422 | } | |
2423 | ||
7c673cae FG |
2424 | // cluster snapshot? |
2425 | if (inc.cluster_snapshot.length()) { | |
2426 | cluster_snapshot = inc.cluster_snapshot; | |
2427 | cluster_snapshot_epoch = inc.epoch; | |
2428 | } else { | |
2429 | cluster_snapshot.clear(); | |
2430 | cluster_snapshot_epoch = 0; | |
2431 | } | |
2432 | ||
2433 | if (inc.new_nearfull_ratio >= 0) { | |
2434 | nearfull_ratio = inc.new_nearfull_ratio; | |
2435 | } | |
2436 | if (inc.new_backfillfull_ratio >= 0) { | |
2437 | backfillfull_ratio = inc.new_backfillfull_ratio; | |
2438 | } | |
2439 | if (inc.new_full_ratio >= 0) { | |
2440 | full_ratio = inc.new_full_ratio; | |
2441 | } | |
9f95a23c | 2442 | if (inc.new_require_min_compat_client > ceph_release_t::unknown) { |
7c673cae FG |
2443 | require_min_compat_client = inc.new_require_min_compat_client; |
2444 | } | |
9f95a23c | 2445 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
31f18b77 | 2446 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2447 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 2448 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 2449 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
2450 | } |
2451 | } | |
7c673cae | 2452 | |
9f95a23c | 2453 | if (inc.new_require_osd_release >= ceph_release_t::unknown) { |
11fdf7f2 | 2454 | require_osd_release = inc.new_require_osd_release; |
9f95a23c | 2455 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
2456 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
2457 | } | |
2458 | } | |
7c673cae FG |
2459 | // do new crush map last (after up/down stuff) |
2460 | if (inc.crush.length()) { | |
9f95a23c | 2461 | ceph::buffer::list bl(inc.crush); |
11fdf7f2 | 2462 | auto blp = bl.cbegin(); |
7c673cae FG |
2463 | crush.reset(new CrushWrapper); |
2464 | crush->decode(blp); | |
9f95a23c | 2465 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 FG |
2466 | // only increment if this is a luminous-encoded osdmap, lest |
2467 | // the mon's crush_version diverge from what the osds or others | |
2468 | // are decoding and applying on their end. if we won't encode | |
2469 | // it in the canonical version, don't change it. | |
2470 | ++crush_version; | |
2471 | } | |
81eedcae TL |
2472 | for (auto it = device_class_flags.begin(); |
2473 | it != device_class_flags.end();) { | |
2474 | const char* class_name = crush->get_class_name(it->first); | |
2475 | if (!class_name) // device class is gone | |
2476 | it = device_class_flags.erase(it); | |
2477 | else | |
2478 | it++; | |
2479 | } | |
7c673cae FG |
2480 | } |
2481 | ||
f67539c2 TL |
2482 | if (inc.change_stretch_mode) { |
2483 | stretch_mode_enabled = inc.stretch_mode_enabled; | |
2484 | stretch_bucket_count = inc.new_stretch_bucket_count; | |
2485 | degraded_stretch_mode = inc.new_degraded_stretch_mode; | |
2486 | recovering_stretch_mode = inc.new_recovering_stretch_mode; | |
2487 | stretch_mode_bucket = inc.new_stretch_mode_bucket; | |
2488 | } | |
2489 | ||
7c673cae FG |
2490 | calc_num_osds(); |
2491 | _calc_up_osd_features(); | |
2492 | return 0; | |
2493 | } | |
2494 | ||
2495 | // mapping | |
2496 | int OSDMap::map_to_pg( | |
2497 | int64_t poolid, | |
2498 | const string& name, | |
2499 | const string& key, | |
2500 | const string& nspace, | |
2501 | pg_t *pg) const | |
2502 | { | |
2503 | // calculate ps (placement seed) | |
2504 | const pg_pool_t *pool = get_pg_pool(poolid); | |
2505 | if (!pool) | |
2506 | return -ENOENT; | |
2507 | ps_t ps; | |
2508 | if (!key.empty()) | |
2509 | ps = pool->hash_key(key, nspace); | |
2510 | else | |
2511 | ps = pool->hash_key(name, nspace); | |
2512 | *pg = pg_t(ps, poolid); | |
2513 | return 0; | |
2514 | } | |
2515 | ||
2516 | int OSDMap::object_locator_to_pg( | |
2517 | const object_t& oid, const object_locator_t& loc, pg_t &pg) const | |
2518 | { | |
2519 | if (loc.hash >= 0) { | |
2520 | if (!get_pg_pool(loc.get_pool())) { | |
2521 | return -ENOENT; | |
2522 | } | |
2523 | pg = pg_t(loc.hash, loc.get_pool()); | |
2524 | return 0; | |
2525 | } | |
2526 | return map_to_pg(loc.get_pool(), oid.name, loc.key, loc.nspace, &pg); | |
2527 | } | |
2528 | ||
2529 | ceph_object_layout OSDMap::make_object_layout( | |
2530 | object_t oid, int pg_pool, string nspace) const | |
2531 | { | |
2532 | object_locator_t loc(pg_pool, nspace); | |
2533 | ||
2534 | ceph_object_layout ol; | |
2535 | pg_t pgid = object_locator_to_pg(oid, loc); | |
2536 | ol.ol_pgid = pgid.get_old_pg().v; | |
2537 | ol.ol_stripe_unit = 0; | |
2538 | return ol; | |
2539 | } | |
2540 | ||
2541 | void OSDMap::_remove_nonexistent_osds(const pg_pool_t& pool, | |
2542 | vector<int>& osds) const | |
2543 | { | |
2544 | if (pool.can_shift_osds()) { | |
2545 | unsigned removed = 0; | |
2546 | for (unsigned i = 0; i < osds.size(); i++) { | |
2547 | if (!exists(osds[i])) { | |
2548 | removed++; | |
2549 | continue; | |
2550 | } | |
2551 | if (removed) { | |
2552 | osds[i - removed] = osds[i]; | |
2553 | } | |
2554 | } | |
2555 | if (removed) | |
2556 | osds.resize(osds.size() - removed); | |
2557 | } else { | |
2558 | for (auto& osd : osds) { | |
2559 | if (!exists(osd)) | |
2560 | osd = CRUSH_ITEM_NONE; | |
2561 | } | |
2562 | } | |
2563 | } | |
2564 | ||
31f18b77 | 2565 | void OSDMap::_pg_to_raw_osds( |
7c673cae FG |
2566 | const pg_pool_t& pool, pg_t pg, |
2567 | vector<int> *osds, | |
2568 | ps_t *ppps) const | |
2569 | { | |
2570 | // map to osds[] | |
2571 | ps_t pps = pool.raw_pg_to_pps(pg); // placement ps | |
2572 | unsigned size = pool.get_size(); | |
2573 | ||
2574 | // what crush rule? | |
20effc67 | 2575 | int ruleno = pool.get_crush_rule(); |
7c673cae FG |
2576 | if (ruleno >= 0) |
2577 | crush->do_rule(ruleno, pps, *osds, size, osd_weight, pg.pool()); | |
2578 | ||
2579 | _remove_nonexistent_osds(pool, *osds); | |
2580 | ||
2581 | if (ppps) | |
2582 | *ppps = pps; | |
7c673cae FG |
2583 | } |
2584 | ||
2585 | int OSDMap::_pick_primary(const vector<int>& osds) const | |
2586 | { | |
2587 | for (auto osd : osds) { | |
2588 | if (osd != CRUSH_ITEM_NONE) { | |
2589 | return osd; | |
2590 | } | |
2591 | } | |
2592 | return -1; | |
2593 | } | |
2594 | ||
224ce89b | 2595 | void OSDMap::_apply_upmap(const pg_pool_t& pi, pg_t raw_pg, vector<int> *raw) const |
7c673cae FG |
2596 | { |
2597 | pg_t pg = pi.raw_pg_to_pg(raw_pg); | |
2598 | auto p = pg_upmap.find(pg); | |
2599 | if (p != pg_upmap.end()) { | |
2600 | // make sure targets aren't marked out | |
2601 | for (auto osd : p->second) { | |
91327a77 AA |
2602 | if (osd != CRUSH_ITEM_NONE && osd < max_osd && osd >= 0 && |
2603 | osd_weight[osd] == 0) { | |
7c673cae FG |
2604 | // reject/ignore the explicit mapping |
2605 | return; | |
2606 | } | |
2607 | } | |
2608 | *raw = vector<int>(p->second.begin(), p->second.end()); | |
224ce89b | 2609 | // continue to check and apply pg_upmap_items if any |
7c673cae FG |
2610 | } |
2611 | ||
2612 | auto q = pg_upmap_items.find(pg); | |
2613 | if (q != pg_upmap_items.end()) { | |
181888fb FG |
2614 | // NOTE: this approach does not allow a bidirectional swap, |
2615 | // e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1]. | |
2616 | for (auto& r : q->second) { | |
2617 | // make sure the replacement value doesn't already appear | |
2618 | bool exists = false; | |
2619 | ssize_t pos = -1; | |
2620 | for (unsigned i = 0; i < raw->size(); ++i) { | |
2621 | int osd = (*raw)[i]; | |
2622 | if (osd == r.second) { | |
2623 | exists = true; | |
2624 | break; | |
2625 | } | |
2626 | // ignore mapping if target is marked out (or invalid osd id) | |
2627 | if (osd == r.first && | |
2628 | pos < 0 && | |
2629 | !(r.second != CRUSH_ITEM_NONE && r.second < max_osd && | |
91327a77 | 2630 | r.second >= 0 && osd_weight[r.second] == 0)) { |
181888fb FG |
2631 | pos = i; |
2632 | } | |
2633 | } | |
2634 | if (!exists && pos >= 0) { | |
2635 | (*raw)[pos] = r.second; | |
7c673cae FG |
2636 | } |
2637 | } | |
2638 | } | |
2639 | } | |
2640 | ||
2641 | // pg -> (up osd list) | |
2642 | void OSDMap::_raw_to_up_osds(const pg_pool_t& pool, const vector<int>& raw, | |
2643 | vector<int> *up) const | |
2644 | { | |
2645 | if (pool.can_shift_osds()) { | |
2646 | // shift left | |
2647 | up->clear(); | |
2648 | up->reserve(raw.size()); | |
2649 | for (unsigned i=0; i<raw.size(); i++) { | |
2650 | if (!exists(raw[i]) || is_down(raw[i])) | |
2651 | continue; | |
2652 | up->push_back(raw[i]); | |
2653 | } | |
2654 | } else { | |
2655 | // set down/dne devices to NONE | |
2656 | up->resize(raw.size()); | |
2657 | for (int i = raw.size() - 1; i >= 0; --i) { | |
2658 | if (!exists(raw[i]) || is_down(raw[i])) { | |
2659 | (*up)[i] = CRUSH_ITEM_NONE; | |
2660 | } else { | |
2661 | (*up)[i] = raw[i]; | |
2662 | } | |
2663 | } | |
2664 | } | |
2665 | } | |
2666 | ||
2667 | void OSDMap::_apply_primary_affinity(ps_t seed, | |
2668 | const pg_pool_t& pool, | |
2669 | vector<int> *osds, | |
2670 | int *primary) const | |
2671 | { | |
2672 | // do we have any non-default primary_affinity values for these osds? | |
2673 | if (!osd_primary_affinity) | |
2674 | return; | |
2675 | ||
2676 | bool any = false; | |
2677 | for (const auto osd : *osds) { | |
2678 | if (osd != CRUSH_ITEM_NONE && | |
2679 | (*osd_primary_affinity)[osd] != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
2680 | any = true; | |
2681 | break; | |
2682 | } | |
2683 | } | |
2684 | if (!any) | |
2685 | return; | |
2686 | ||
2687 | // pick the primary. feed both the seed (for the pg) and the osd | |
2688 | // into the hash/rng so that a proportional fraction of an osd's pgs | |
2689 | // get rejected as primary. | |
2690 | int pos = -1; | |
2691 | for (unsigned i = 0; i < osds->size(); ++i) { | |
2692 | int o = (*osds)[i]; | |
2693 | if (o == CRUSH_ITEM_NONE) | |
2694 | continue; | |
2695 | unsigned a = (*osd_primary_affinity)[o]; | |
2696 | if (a < CEPH_OSD_MAX_PRIMARY_AFFINITY && | |
2697 | (crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
2698 | seed, o) >> 16) >= a) { | |
2699 | // we chose not to use this primary. note it anyway as a | |
2700 | // fallback in case we don't pick anyone else, but keep looking. | |
2701 | if (pos < 0) | |
2702 | pos = i; | |
2703 | } else { | |
2704 | pos = i; | |
2705 | break; | |
2706 | } | |
2707 | } | |
2708 | if (pos < 0) | |
2709 | return; | |
2710 | ||
2711 | *primary = (*osds)[pos]; | |
2712 | ||
2713 | if (pool.can_shift_osds() && pos > 0) { | |
2714 | // move the new primary to the front. | |
2715 | for (int i = pos; i > 0; --i) { | |
2716 | (*osds)[i] = (*osds)[i-1]; | |
2717 | } | |
2718 | (*osds)[0] = *primary; | |
2719 | } | |
2720 | } | |
2721 | ||
2722 | void OSDMap::_get_temp_osds(const pg_pool_t& pool, pg_t pg, | |
2723 | vector<int> *temp_pg, int *temp_primary) const | |
2724 | { | |
2725 | pg = pool.raw_pg_to_pg(pg); | |
2726 | const auto p = pg_temp->find(pg); | |
2727 | temp_pg->clear(); | |
2728 | if (p != pg_temp->end()) { | |
2729 | for (unsigned i=0; i<p->second.size(); i++) { | |
2730 | if (!exists(p->second[i]) || is_down(p->second[i])) { | |
2731 | if (pool.can_shift_osds()) { | |
2732 | continue; | |
2733 | } else { | |
2734 | temp_pg->push_back(CRUSH_ITEM_NONE); | |
2735 | } | |
2736 | } else { | |
2737 | temp_pg->push_back(p->second[i]); | |
2738 | } | |
2739 | } | |
2740 | } | |
2741 | const auto &pp = primary_temp->find(pg); | |
2742 | *temp_primary = -1; | |
2743 | if (pp != primary_temp->end()) { | |
2744 | *temp_primary = pp->second; | |
2745 | } else if (!temp_pg->empty()) { // apply pg_temp's primary | |
2746 | for (unsigned i = 0; i < temp_pg->size(); ++i) { | |
2747 | if ((*temp_pg)[i] != CRUSH_ITEM_NONE) { | |
2748 | *temp_primary = (*temp_pg)[i]; | |
2749 | break; | |
2750 | } | |
2751 | } | |
2752 | } | |
2753 | } | |
2754 | ||
31f18b77 | 2755 | void OSDMap::pg_to_raw_osds(pg_t pg, vector<int> *raw, int *primary) const |
7c673cae | 2756 | { |
7c673cae | 2757 | const pg_pool_t *pool = get_pg_pool(pg.pool()); |
11fdf7f2 TL |
2758 | if (!pool) { |
2759 | *primary = -1; | |
2760 | raw->clear(); | |
31f18b77 | 2761 | return; |
11fdf7f2 | 2762 | } |
31f18b77 | 2763 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
11fdf7f2 | 2764 | *primary = _pick_primary(*raw); |
7c673cae FG |
2765 | } |
2766 | ||
494da23a TL |
2767 | void OSDMap::pg_to_raw_upmap(pg_t pg, vector<int>*raw, |
2768 | vector<int> *raw_upmap) const | |
a8e16298 TL |
2769 | { |
2770 | auto pool = get_pg_pool(pg.pool()); | |
2771 | if (!pool) { | |
2772 | raw_upmap->clear(); | |
2773 | return; | |
2774 | } | |
494da23a TL |
2775 | _pg_to_raw_osds(*pool, pg, raw, NULL); |
2776 | *raw_upmap = *raw; | |
a8e16298 TL |
2777 | _apply_upmap(*pool, pg, raw_upmap); |
2778 | } | |
2779 | ||
7c673cae FG |
2780 | void OSDMap::pg_to_raw_up(pg_t pg, vector<int> *up, int *primary) const |
2781 | { | |
2782 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2783 | if (!pool) { | |
11fdf7f2 TL |
2784 | *primary = -1; |
2785 | up->clear(); | |
7c673cae FG |
2786 | return; |
2787 | } | |
2788 | vector<int> raw; | |
2789 | ps_t pps; | |
2790 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2791 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2792 | _raw_to_up_osds(*pool, raw, up); |
2793 | *primary = _pick_primary(raw); | |
2794 | _apply_primary_affinity(pps, *pool, up, primary); | |
2795 | } | |
31f18b77 | 2796 | |
7c673cae FG |
2797 | void OSDMap::_pg_to_up_acting_osds( |
2798 | const pg_t& pg, vector<int> *up, int *up_primary, | |
2799 | vector<int> *acting, int *acting_primary, | |
2800 | bool raw_pg_to_pg) const | |
2801 | { | |
2802 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
2803 | if (!pool || | |
2804 | (!raw_pg_to_pg && pg.ps() >= pool->get_pg_num())) { | |
2805 | if (up) | |
2806 | up->clear(); | |
2807 | if (up_primary) | |
2808 | *up_primary = -1; | |
2809 | if (acting) | |
2810 | acting->clear(); | |
2811 | if (acting_primary) | |
2812 | *acting_primary = -1; | |
2813 | return; | |
2814 | } | |
2815 | vector<int> raw; | |
2816 | vector<int> _up; | |
2817 | vector<int> _acting; | |
2818 | int _up_primary; | |
2819 | int _acting_primary; | |
2820 | ps_t pps; | |
2821 | _get_temp_osds(*pool, pg, &_acting, &_acting_primary); | |
2822 | if (_acting.empty() || up || up_primary) { | |
2823 | _pg_to_raw_osds(*pool, pg, &raw, &pps); | |
224ce89b | 2824 | _apply_upmap(*pool, pg, &raw); |
7c673cae FG |
2825 | _raw_to_up_osds(*pool, raw, &_up); |
2826 | _up_primary = _pick_primary(_up); | |
2827 | _apply_primary_affinity(pps, *pool, &_up, &_up_primary); | |
2828 | if (_acting.empty()) { | |
2829 | _acting = _up; | |
2830 | if (_acting_primary == -1) { | |
2831 | _acting_primary = _up_primary; | |
2832 | } | |
2833 | } | |
2834 | ||
2835 | if (up) | |
2836 | up->swap(_up); | |
2837 | if (up_primary) | |
2838 | *up_primary = _up_primary; | |
2839 | } | |
2840 | ||
2841 | if (acting) | |
2842 | acting->swap(_acting); | |
2843 | if (acting_primary) | |
2844 | *acting_primary = _acting_primary; | |
2845 | } | |
2846 | ||
9f95a23c | 2847 | int OSDMap::calc_pg_role_broken(int osd, const vector<int>& acting, int nrep) |
7c673cae | 2848 | { |
9f95a23c TL |
2849 | // This implementation is broken for EC PGs since the osd may appear |
2850 | // multiple times in the acting set. See | |
2851 | // https://tracker.ceph.com/issues/43213 | |
7c673cae FG |
2852 | if (!nrep) |
2853 | nrep = acting.size(); | |
2854 | for (int i=0; i<nrep; i++) | |
2855 | if (acting[i] == osd) | |
2856 | return i; | |
2857 | return -1; | |
2858 | } | |
2859 | ||
9f95a23c | 2860 | int OSDMap::calc_pg_role(pg_shard_t who, const vector<int>& acting) |
7c673cae | 2861 | { |
9f95a23c TL |
2862 | int nrep = acting.size(); |
2863 | if (who.shard == shard_id_t::NO_SHARD) { | |
2864 | for (int i=0; i<nrep; i++) { | |
2865 | if (acting[i] == who.osd) { | |
2866 | return i; | |
2867 | } | |
2868 | } | |
2869 | } else { | |
2870 | if (who.shard < nrep && acting[who.shard] == who.osd) { | |
2871 | return who.shard; | |
2872 | } | |
2873 | } | |
2874 | return -1; | |
7c673cae FG |
2875 | } |
2876 | ||
9f95a23c | 2877 | bool OSDMap::primary_changed_broken( |
7c673cae FG |
2878 | int oldprimary, |
2879 | const vector<int> &oldacting, | |
2880 | int newprimary, | |
2881 | const vector<int> &newacting) | |
2882 | { | |
2883 | if (oldacting.empty() && newacting.empty()) | |
2884 | return false; // both still empty | |
2885 | if (oldacting.empty() ^ newacting.empty()) | |
2886 | return true; // was empty, now not, or vice versa | |
2887 | if (oldprimary != newprimary) | |
2888 | return true; // primary changed | |
9f95a23c TL |
2889 | if (calc_pg_role_broken(oldprimary, oldacting) != |
2890 | calc_pg_role_broken(newprimary, newacting)) | |
7c673cae FG |
2891 | return true; |
2892 | return false; // same primary (tho replicas may have changed) | |
2893 | } | |
2894 | ||
28e407b8 AA |
2895 | uint64_t OSDMap::get_encoding_features() const |
2896 | { | |
2897 | uint64_t f = SIGNIFICANT_FEATURES; | |
9f95a23c TL |
2898 | if (require_osd_release < ceph_release_t::octopus) { |
2899 | f &= ~CEPH_FEATURE_SERVER_OCTOPUS; | |
2900 | } | |
2901 | if (require_osd_release < ceph_release_t::nautilus) { | |
11fdf7f2 TL |
2902 | f &= ~CEPH_FEATURE_SERVER_NAUTILUS; |
2903 | } | |
9f95a23c | 2904 | if (require_osd_release < ceph_release_t::mimic) { |
11fdf7f2 TL |
2905 | f &= ~CEPH_FEATURE_SERVER_MIMIC; |
2906 | } | |
9f95a23c | 2907 | if (require_osd_release < ceph_release_t::luminous) { |
28e407b8 AA |
2908 | f &= ~(CEPH_FEATURE_SERVER_LUMINOUS | |
2909 | CEPH_FEATURE_CRUSH_CHOOSE_ARGS); | |
2910 | } | |
9f95a23c | 2911 | if (require_osd_release < ceph_release_t::kraken) { |
28e407b8 | 2912 | f &= ~(CEPH_FEATURE_SERVER_KRAKEN | |
1adf2230 | 2913 | CEPH_FEATURE_MSG_ADDR2); |
28e407b8 | 2914 | } |
9f95a23c | 2915 | if (require_osd_release < ceph_release_t::jewel) { |
28e407b8 | 2916 | f &= ~(CEPH_FEATURE_SERVER_JEWEL | |
1adf2230 AA |
2917 | CEPH_FEATURE_NEW_OSDOP_ENCODING | |
2918 | CEPH_FEATURE_CRUSH_TUNABLES5); | |
28e407b8 AA |
2919 | } |
2920 | return f; | |
2921 | } | |
7c673cae FG |
2922 | |
2923 | // serialize, unserialize | |
9f95a23c | 2924 | void OSDMap::encode_client_old(ceph::buffer::list& bl) const |
7c673cae | 2925 | { |
11fdf7f2 | 2926 | using ceph::encode; |
7c673cae | 2927 | __u16 v = 5; |
11fdf7f2 | 2928 | encode(v, bl); |
7c673cae FG |
2929 | |
2930 | // base | |
11fdf7f2 TL |
2931 | encode(fsid, bl); |
2932 | encode(epoch, bl); | |
2933 | encode(created, bl); | |
2934 | encode(modified, bl); | |
7c673cae | 2935 | |
11fdf7f2 | 2936 | // for encode(pools, bl); |
7c673cae | 2937 | __u32 n = pools.size(); |
11fdf7f2 | 2938 | encode(n, bl); |
7c673cae FG |
2939 | |
2940 | for (const auto &pool : pools) { | |
2941 | n = pool.first; | |
11fdf7f2 TL |
2942 | encode(n, bl); |
2943 | encode(pool.second, bl, 0); | |
7c673cae | 2944 | } |
11fdf7f2 | 2945 | // for encode(pool_name, bl); |
7c673cae | 2946 | n = pool_name.size(); |
11fdf7f2 | 2947 | encode(n, bl); |
7c673cae FG |
2948 | for (const auto &pname : pool_name) { |
2949 | n = pname.first; | |
11fdf7f2 TL |
2950 | encode(n, bl); |
2951 | encode(pname.second, bl); | |
7c673cae | 2952 | } |
11fdf7f2 | 2953 | // for encode(pool_max, bl); |
7c673cae | 2954 | n = pool_max; |
11fdf7f2 | 2955 | encode(n, bl); |
7c673cae | 2956 | |
11fdf7f2 | 2957 | encode(flags, bl); |
7c673cae | 2958 | |
11fdf7f2 | 2959 | encode(max_osd, bl); |
31f18b77 FG |
2960 | { |
2961 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 2962 | encode(n, bl); |
31f18b77 | 2963 | for (auto s : osd_state) { |
11fdf7f2 | 2964 | encode((uint8_t)s, bl); |
31f18b77 FG |
2965 | } |
2966 | } | |
11fdf7f2 TL |
2967 | encode(osd_weight, bl); |
2968 | encode(osd_addrs->client_addrs, bl, 0); | |
7c673cae | 2969 | |
11fdf7f2 | 2970 | // for encode(pg_temp, bl); |
7c673cae | 2971 | n = pg_temp->size(); |
11fdf7f2 | 2972 | encode(n, bl); |
f67539c2 | 2973 | for (const auto& pg : *pg_temp) { |
7c673cae | 2974 | old_pg_t opg = pg.first.get_old_pg(); |
11fdf7f2 TL |
2975 | encode(opg, bl); |
2976 | encode(pg.second, bl); | |
7c673cae FG |
2977 | } |
2978 | ||
2979 | // crush | |
9f95a23c | 2980 | ceph::buffer::list cbl; |
7c673cae | 2981 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 2982 | encode(cbl, bl); |
7c673cae FG |
2983 | } |
2984 | ||
9f95a23c | 2985 | void OSDMap::encode_classic(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 2986 | { |
11fdf7f2 | 2987 | using ceph::encode; |
7c673cae FG |
2988 | if ((features & CEPH_FEATURE_PGID64) == 0) { |
2989 | encode_client_old(bl); | |
2990 | return; | |
2991 | } | |
2992 | ||
2993 | __u16 v = 6; | |
11fdf7f2 | 2994 | encode(v, bl); |
7c673cae FG |
2995 | |
2996 | // base | |
11fdf7f2 TL |
2997 | encode(fsid, bl); |
2998 | encode(epoch, bl); | |
2999 | encode(created, bl); | |
3000 | encode(modified, bl); | |
7c673cae | 3001 | |
11fdf7f2 TL |
3002 | encode(pools, bl, features); |
3003 | encode(pool_name, bl); | |
3004 | encode(pool_max, bl); | |
7c673cae | 3005 | |
11fdf7f2 | 3006 | encode(flags, bl); |
7c673cae | 3007 | |
11fdf7f2 | 3008 | encode(max_osd, bl); |
31f18b77 FG |
3009 | { |
3010 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 3011 | encode(n, bl); |
31f18b77 | 3012 | for (auto s : osd_state) { |
11fdf7f2 | 3013 | encode((uint8_t)s, bl); |
31f18b77 FG |
3014 | } |
3015 | } | |
11fdf7f2 TL |
3016 | encode(osd_weight, bl); |
3017 | encode(osd_addrs->client_addrs, bl, features); | |
7c673cae | 3018 | |
11fdf7f2 | 3019 | encode(*pg_temp, bl); |
7c673cae FG |
3020 | |
3021 | // crush | |
9f95a23c | 3022 | ceph::buffer::list cbl; |
7c673cae | 3023 | crush->encode(cbl, 0 /* legacy (no) features */); |
11fdf7f2 | 3024 | encode(cbl, bl); |
7c673cae FG |
3025 | |
3026 | // extended | |
3027 | __u16 ev = 10; | |
11fdf7f2 TL |
3028 | encode(ev, bl); |
3029 | encode(osd_addrs->hb_back_addrs, bl, features); | |
3030 | encode(osd_info, bl); | |
f67539c2 | 3031 | encode(blocklist, bl, features); |
11fdf7f2 TL |
3032 | encode(osd_addrs->cluster_addrs, bl, features); |
3033 | encode(cluster_snapshot_epoch, bl); | |
3034 | encode(cluster_snapshot, bl); | |
3035 | encode(*osd_uuid, bl); | |
9f95a23c | 3036 | encode(osd_xinfo, bl, features); |
11fdf7f2 | 3037 | encode(osd_addrs->hb_front_addrs, bl, features); |
7c673cae FG |
3038 | } |
3039 | ||
11fdf7f2 TL |
3040 | /* for a description of osdmap versions, and when they were introduced, please |
3041 | * refer to | |
3042 | * doc/dev/osd_internals/osdmap_versions.txt | |
3043 | */ | |
9f95a23c | 3044 | void OSDMap::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 3045 | { |
11fdf7f2 | 3046 | using ceph::encode; |
7c673cae FG |
3047 | if ((features & CEPH_FEATURE_OSDMAP_ENC) == 0) { |
3048 | encode_classic(bl, features); | |
3049 | return; | |
3050 | } | |
3051 | ||
3052 | // only a select set of callers should *ever* be encoding new | |
3053 | // OSDMaps. others should be passing around the canonical encoded | |
3054 | // buffers from on high. select out those callers by passing in an | |
3055 | // "impossible" feature bit. | |
11fdf7f2 | 3056 | ceph_assert(features & CEPH_FEATURE_RESERVED); |
7c673cae FG |
3057 | features &= ~CEPH_FEATURE_RESERVED; |
3058 | ||
3059 | size_t start_offset = bl.length(); | |
3060 | size_t tail_offset; | |
11fdf7f2 | 3061 | size_t crc_offset; |
9f95a23c | 3062 | std::optional<ceph::buffer::list::contiguous_filler> crc_filler; |
7c673cae FG |
3063 | |
3064 | // meta-encoding: how we include client-used and osd-specific data | |
3065 | ENCODE_START(8, 7, bl); | |
3066 | ||
3067 | { | |
28e407b8 AA |
3068 | // NOTE: any new encoding dependencies must be reflected by |
3069 | // SIGNIFICANT_FEATURES | |
11fdf7f2 | 3070 | uint8_t v = 9; |
31f18b77 | 3071 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
7c673cae | 3072 | v = 3; |
11fdf7f2 TL |
3073 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
3074 | v = 6; | |
3075 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
3076 | v = 7; | |
7c673cae FG |
3077 | } |
3078 | ENCODE_START(v, 1, bl); // client-usable data | |
3079 | // base | |
11fdf7f2 TL |
3080 | encode(fsid, bl); |
3081 | encode(epoch, bl); | |
3082 | encode(created, bl); | |
3083 | encode(modified, bl); | |
7c673cae | 3084 | |
11fdf7f2 TL |
3085 | encode(pools, bl, features); |
3086 | encode(pool_name, bl); | |
3087 | encode(pool_max, bl); | |
7c673cae | 3088 | |
31f18b77 FG |
3089 | if (v < 4) { |
3090 | decltype(flags) f = flags; | |
9f95a23c | 3091 | if (require_osd_release >= ceph_release_t::luminous) |
c07f9fc5 | 3092 | f |= CEPH_OSDMAP_REQUIRE_LUMINOUS | CEPH_OSDMAP_RECOVERY_DELETES; |
9f95a23c | 3093 | else if (require_osd_release == ceph_release_t::kraken) |
31f18b77 | 3094 | f |= CEPH_OSDMAP_REQUIRE_KRAKEN; |
9f95a23c | 3095 | else if (require_osd_release == ceph_release_t::jewel) |
31f18b77 | 3096 | f |= CEPH_OSDMAP_REQUIRE_JEWEL; |
11fdf7f2 | 3097 | encode(f, bl); |
31f18b77 | 3098 | } else { |
11fdf7f2 | 3099 | encode(flags, bl); |
31f18b77 | 3100 | } |
7c673cae | 3101 | |
11fdf7f2 | 3102 | encode(max_osd, bl); |
31f18b77 | 3103 | if (v >= 5) { |
11fdf7f2 | 3104 | encode(osd_state, bl); |
31f18b77 FG |
3105 | } else { |
3106 | uint32_t n = osd_state.size(); | |
11fdf7f2 | 3107 | encode(n, bl); |
31f18b77 | 3108 | for (auto s : osd_state) { |
11fdf7f2 | 3109 | encode((uint8_t)s, bl); |
31f18b77 FG |
3110 | } |
3111 | } | |
11fdf7f2 TL |
3112 | encode(osd_weight, bl); |
3113 | if (v >= 8) { | |
3114 | encode(osd_addrs->client_addrs, bl, features); | |
3115 | } else { | |
3116 | encode_addrvec_pvec_as_addr(osd_addrs->client_addrs, bl, features); | |
3117 | } | |
7c673cae | 3118 | |
11fdf7f2 TL |
3119 | encode(*pg_temp, bl); |
3120 | encode(*primary_temp, bl); | |
7c673cae | 3121 | if (osd_primary_affinity) { |
11fdf7f2 | 3122 | encode(*osd_primary_affinity, bl); |
7c673cae FG |
3123 | } else { |
3124 | vector<__u32> v; | |
11fdf7f2 | 3125 | encode(v, bl); |
7c673cae FG |
3126 | } |
3127 | ||
3128 | // crush | |
9f95a23c | 3129 | ceph::buffer::list cbl; |
7c673cae | 3130 | crush->encode(cbl, features); |
11fdf7f2 TL |
3131 | encode(cbl, bl); |
3132 | encode(erasure_code_profiles, bl); | |
7c673cae FG |
3133 | |
3134 | if (v >= 4) { | |
11fdf7f2 TL |
3135 | encode(pg_upmap, bl); |
3136 | encode(pg_upmap_items, bl); | |
7c673cae | 3137 | } else { |
11fdf7f2 TL |
3138 | ceph_assert(pg_upmap.empty()); |
3139 | ceph_assert(pg_upmap_items.empty()); | |
7c673cae | 3140 | } |
31f18b77 | 3141 | if (v >= 6) { |
11fdf7f2 TL |
3142 | encode(crush_version, bl); |
3143 | } | |
3144 | if (v >= 7) { | |
3145 | encode(new_removed_snaps, bl); | |
3146 | encode(new_purged_snaps, bl); | |
3147 | } | |
3148 | if (v >= 9) { | |
3149 | encode(last_up_change, bl); | |
3150 | encode(last_in_change, bl); | |
31f18b77 | 3151 | } |
7c673cae FG |
3152 | ENCODE_FINISH(bl); // client-usable data |
3153 | } | |
3154 | ||
3155 | { | |
28e407b8 AA |
3156 | // NOTE: any new encoding dependencies must be reflected by |
3157 | // SIGNIFICANT_FEATURES | |
33c7a0ef | 3158 | uint8_t target_v = 9; // when bumping this, be aware of range blocklist |
7c673cae FG |
3159 | if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
3160 | target_v = 1; | |
11fdf7f2 TL |
3161 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
3162 | target_v = 5; | |
3163 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
3164 | target_v = 6; | |
7c673cae | 3165 | } |
f67539c2 TL |
3166 | if (stretch_mode_enabled) { |
3167 | target_v = std::max((uint8_t)10, target_v); | |
3168 | } | |
33c7a0ef TL |
3169 | if (!range_blocklist.empty()) { |
3170 | target_v = std::max((uint8_t)11, target_v); | |
3171 | } | |
7c673cae | 3172 | ENCODE_START(target_v, 1, bl); // extended, osd-only data |
11fdf7f2 TL |
3173 | if (target_v < 7) { |
3174 | encode_addrvec_pvec_as_addr(osd_addrs->hb_back_addrs, bl, features); | |
3175 | } else { | |
3176 | encode(osd_addrs->hb_back_addrs, bl, features); | |
3177 | } | |
3178 | encode(osd_info, bl); | |
7c673cae FG |
3179 | { |
3180 | // put this in a sorted, ordered map<> so that we encode in a | |
3181 | // deterministic order. | |
f67539c2 TL |
3182 | map<entity_addr_t,utime_t> blocklist_map; |
3183 | for (const auto &addr : blocklist) | |
3184 | blocklist_map.insert(make_pair(addr.first, addr.second)); | |
3185 | encode(blocklist_map, bl, features); | |
11fdf7f2 TL |
3186 | } |
3187 | if (target_v < 7) { | |
3188 | encode_addrvec_pvec_as_addr(osd_addrs->cluster_addrs, bl, features); | |
3189 | } else { | |
3190 | encode(osd_addrs->cluster_addrs, bl, features); | |
3191 | } | |
3192 | encode(cluster_snapshot_epoch, bl); | |
3193 | encode(cluster_snapshot, bl); | |
3194 | encode(*osd_uuid, bl); | |
9f95a23c | 3195 | encode(osd_xinfo, bl, features); |
11fdf7f2 TL |
3196 | if (target_v < 7) { |
3197 | encode_addrvec_pvec_as_addr(osd_addrs->hb_front_addrs, bl, features); | |
3198 | } else { | |
3199 | encode(osd_addrs->hb_front_addrs, bl, features); | |
3200 | } | |
7c673cae | 3201 | if (target_v >= 2) { |
11fdf7f2 TL |
3202 | encode(nearfull_ratio, bl); |
3203 | encode(full_ratio, bl); | |
3204 | encode(backfillfull_ratio, bl); | |
31f18b77 FG |
3205 | } |
3206 | // 4 was string-based new_require_min_compat_client | |
3207 | if (target_v >= 5) { | |
11fdf7f2 TL |
3208 | encode(require_min_compat_client, bl); |
3209 | encode(require_osd_release, bl); | |
3210 | } | |
3211 | if (target_v >= 6) { | |
3212 | encode(removed_snaps_queue, bl); | |
7c673cae | 3213 | } |
81eedcae TL |
3214 | if (target_v >= 8) { |
3215 | encode(crush_node_flags, bl); | |
3216 | } | |
3217 | if (target_v >= 9) { | |
3218 | encode(device_class_flags, bl); | |
3219 | } | |
f67539c2 TL |
3220 | if (target_v >= 10) { |
3221 | encode(stretch_mode_enabled, bl); | |
3222 | encode(stretch_bucket_count, bl); | |
3223 | encode(degraded_stretch_mode, bl); | |
3224 | encode(recovering_stretch_mode, bl); | |
3225 | encode(stretch_mode_bucket, bl); | |
3226 | } | |
33c7a0ef TL |
3227 | if (target_v >= 11) { |
3228 | ::encode(range_blocklist, bl, features); | |
3229 | } | |
7c673cae FG |
3230 | ENCODE_FINISH(bl); // osd-only data |
3231 | } | |
3232 | ||
11fdf7f2 TL |
3233 | crc_offset = bl.length(); |
3234 | crc_filler = bl.append_hole(sizeof(uint32_t)); | |
7c673cae FG |
3235 | tail_offset = bl.length(); |
3236 | ||
3237 | ENCODE_FINISH(bl); // meta-encoding wrapper | |
3238 | ||
3239 | // fill in crc | |
9f95a23c | 3240 | ceph::buffer::list front; |
11fdf7f2 | 3241 | front.substr_of(bl, start_offset, crc_offset - start_offset); |
7c673cae FG |
3242 | crc = front.crc32c(-1); |
3243 | if (tail_offset < bl.length()) { | |
9f95a23c | 3244 | ceph::buffer::list tail; |
7c673cae FG |
3245 | tail.substr_of(bl, tail_offset, bl.length() - tail_offset); |
3246 | crc = tail.crc32c(crc); | |
3247 | } | |
3248 | ceph_le32 crc_le; | |
3249 | crc_le = crc; | |
11fdf7f2 | 3250 | crc_filler->copy_in(4, (char*)&crc_le); |
7c673cae FG |
3251 | crc_defined = true; |
3252 | } | |
3253 | ||
11fdf7f2 TL |
3254 | /* for a description of osdmap versions, and when they were introduced, please |
3255 | * refer to | |
3256 | * doc/dev/osd_internals/osdmap_versions.txt | |
3257 | */ | |
9f95a23c | 3258 | void OSDMap::decode(ceph::buffer::list& bl) |
7c673cae | 3259 | { |
11fdf7f2 | 3260 | auto p = bl.cbegin(); |
7c673cae FG |
3261 | decode(p); |
3262 | } | |
3263 | ||
9f95a23c | 3264 | void OSDMap::decode_classic(ceph::buffer::list::const_iterator& p) |
7c673cae | 3265 | { |
11fdf7f2 | 3266 | using ceph::decode; |
7c673cae FG |
3267 | __u32 n, t; |
3268 | __u16 v; | |
11fdf7f2 | 3269 | decode(v, p); |
7c673cae FG |
3270 | |
3271 | // base | |
11fdf7f2 TL |
3272 | decode(fsid, p); |
3273 | decode(epoch, p); | |
3274 | decode(created, p); | |
3275 | decode(modified, p); | |
7c673cae FG |
3276 | |
3277 | if (v < 6) { | |
3278 | if (v < 4) { | |
3279 | int32_t max_pools = 0; | |
11fdf7f2 | 3280 | decode(max_pools, p); |
7c673cae FG |
3281 | pool_max = max_pools; |
3282 | } | |
3283 | pools.clear(); | |
11fdf7f2 | 3284 | decode(n, p); |
7c673cae | 3285 | while (n--) { |
11fdf7f2 TL |
3286 | decode(t, p); |
3287 | decode(pools[t], p); | |
7c673cae FG |
3288 | } |
3289 | if (v == 4) { | |
11fdf7f2 | 3290 | decode(n, p); |
7c673cae FG |
3291 | pool_max = n; |
3292 | } else if (v == 5) { | |
3293 | pool_name.clear(); | |
11fdf7f2 | 3294 | decode(n, p); |
7c673cae | 3295 | while (n--) { |
11fdf7f2 TL |
3296 | decode(t, p); |
3297 | decode(pool_name[t], p); | |
7c673cae | 3298 | } |
11fdf7f2 | 3299 | decode(n, p); |
7c673cae FG |
3300 | pool_max = n; |
3301 | } | |
3302 | } else { | |
11fdf7f2 TL |
3303 | decode(pools, p); |
3304 | decode(pool_name, p); | |
3305 | decode(pool_max, p); | |
7c673cae FG |
3306 | } |
3307 | // kludge around some old bug that zeroed out pool_max (#2307) | |
3308 | if (pools.size() && pool_max < pools.rbegin()->first) { | |
3309 | pool_max = pools.rbegin()->first; | |
3310 | } | |
3311 | ||
11fdf7f2 | 3312 | decode(flags, p); |
7c673cae | 3313 | |
11fdf7f2 | 3314 | decode(max_osd, p); |
31f18b77 FG |
3315 | { |
3316 | vector<uint8_t> os; | |
11fdf7f2 | 3317 | decode(os, p); |
31f18b77 FG |
3318 | osd_state.resize(os.size()); |
3319 | for (unsigned i = 0; i < os.size(); ++i) { | |
3320 | osd_state[i] = os[i]; | |
3321 | } | |
3322 | } | |
11fdf7f2 TL |
3323 | decode(osd_weight, p); |
3324 | decode(osd_addrs->client_addrs, p); | |
7c673cae FG |
3325 | if (v <= 5) { |
3326 | pg_temp->clear(); | |
11fdf7f2 | 3327 | decode(n, p); |
7c673cae FG |
3328 | while (n--) { |
3329 | old_pg_t opg; | |
9f95a23c | 3330 | ceph::decode_raw(opg, p); |
31f18b77 | 3331 | mempool::osdmap::vector<int32_t> v; |
11fdf7f2 | 3332 | decode(v, p); |
31f18b77 | 3333 | pg_temp->set(pg_t(opg), v); |
7c673cae FG |
3334 | } |
3335 | } else { | |
11fdf7f2 | 3336 | decode(*pg_temp, p); |
7c673cae FG |
3337 | } |
3338 | ||
3339 | // crush | |
9f95a23c | 3340 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3341 | decode(cbl, p); |
3342 | auto cblp = cbl.cbegin(); | |
7c673cae FG |
3343 | crush->decode(cblp); |
3344 | ||
3345 | // extended | |
3346 | __u16 ev = 0; | |
3347 | if (v >= 5) | |
11fdf7f2 TL |
3348 | decode(ev, p); |
3349 | decode(osd_addrs->hb_back_addrs, p); | |
3350 | decode(osd_info, p); | |
7c673cae | 3351 | if (v < 5) |
11fdf7f2 | 3352 | decode(pool_name, p); |
7c673cae | 3353 | |
f67539c2 | 3354 | decode(blocklist, p); |
7c673cae | 3355 | if (ev >= 6) |
11fdf7f2 | 3356 | decode(osd_addrs->cluster_addrs, p); |
7c673cae | 3357 | else |
11fdf7f2 | 3358 | osd_addrs->cluster_addrs.resize(osd_addrs->client_addrs.size()); |
7c673cae FG |
3359 | |
3360 | if (ev >= 7) { | |
11fdf7f2 TL |
3361 | decode(cluster_snapshot_epoch, p); |
3362 | decode(cluster_snapshot, p); | |
7c673cae FG |
3363 | } |
3364 | ||
3365 | if (ev >= 8) { | |
11fdf7f2 | 3366 | decode(*osd_uuid, p); |
7c673cae FG |
3367 | } else { |
3368 | osd_uuid->resize(max_osd); | |
3369 | } | |
3370 | if (ev >= 9) | |
11fdf7f2 | 3371 | decode(osd_xinfo, p); |
7c673cae FG |
3372 | else |
3373 | osd_xinfo.resize(max_osd); | |
3374 | ||
3375 | if (ev >= 10) | |
11fdf7f2 | 3376 | decode(osd_addrs->hb_front_addrs, p); |
7c673cae | 3377 | else |
11fdf7f2 | 3378 | osd_addrs->hb_front_addrs.resize(osd_addrs->hb_back_addrs.size()); |
7c673cae FG |
3379 | |
3380 | osd_primary_affinity.reset(); | |
3381 | ||
3382 | post_decode(); | |
3383 | } | |
3384 | ||
9f95a23c | 3385 | void OSDMap::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 3386 | { |
11fdf7f2 | 3387 | using ceph::decode; |
7c673cae FG |
3388 | /** |
3389 | * Older encodings of the OSDMap had a single struct_v which | |
3390 | * covered the whole encoding, and was prior to our modern | |
3391 | * stuff which includes a compatv and a size. So if we see | |
3392 | * a struct_v < 7, we must rewind to the beginning and use our | |
3393 | * classic decoder. | |
3394 | */ | |
3395 | size_t start_offset = bl.get_off(); | |
3396 | size_t tail_offset = 0; | |
9f95a23c | 3397 | ceph::buffer::list crc_front, crc_tail; |
7c673cae FG |
3398 | |
3399 | DECODE_START_LEGACY_COMPAT_LEN(8, 7, 7, bl); // wrapper | |
3400 | if (struct_v < 7) { | |
11fdf7f2 | 3401 | bl.seek(start_offset); |
7c673cae FG |
3402 | decode_classic(bl); |
3403 | return; | |
3404 | } | |
3405 | /** | |
3406 | * Since we made it past that hurdle, we can use our normal paths. | |
3407 | */ | |
3408 | { | |
11fdf7f2 | 3409 | DECODE_START(9, bl); // client-usable data |
7c673cae | 3410 | // base |
11fdf7f2 TL |
3411 | decode(fsid, bl); |
3412 | decode(epoch, bl); | |
3413 | decode(created, bl); | |
3414 | decode(modified, bl); | |
7c673cae | 3415 | |
11fdf7f2 TL |
3416 | decode(pools, bl); |
3417 | decode(pool_name, bl); | |
3418 | decode(pool_max, bl); | |
7c673cae | 3419 | |
11fdf7f2 | 3420 | decode(flags, bl); |
7c673cae | 3421 | |
11fdf7f2 | 3422 | decode(max_osd, bl); |
31f18b77 | 3423 | if (struct_v >= 5) { |
11fdf7f2 | 3424 | decode(osd_state, bl); |
31f18b77 FG |
3425 | } else { |
3426 | vector<uint8_t> os; | |
11fdf7f2 | 3427 | decode(os, bl); |
31f18b77 FG |
3428 | osd_state.resize(os.size()); |
3429 | for (unsigned i = 0; i < os.size(); ++i) { | |
3430 | osd_state[i] = os[i]; | |
3431 | } | |
3432 | } | |
11fdf7f2 TL |
3433 | decode(osd_weight, bl); |
3434 | decode(osd_addrs->client_addrs, bl); | |
7c673cae | 3435 | |
11fdf7f2 TL |
3436 | decode(*pg_temp, bl); |
3437 | decode(*primary_temp, bl); | |
3438 | // dates back to firefly. version increased from 2 to 3 still in firefly. | |
3439 | // do we really still need to keep this around? even for old clients? | |
7c673cae FG |
3440 | if (struct_v >= 2) { |
3441 | osd_primary_affinity.reset(new mempool::osdmap::vector<__u32>); | |
11fdf7f2 | 3442 | decode(*osd_primary_affinity, bl); |
7c673cae FG |
3443 | if (osd_primary_affinity->empty()) |
3444 | osd_primary_affinity.reset(); | |
3445 | } else { | |
3446 | osd_primary_affinity.reset(); | |
3447 | } | |
3448 | ||
3449 | // crush | |
9f95a23c | 3450 | ceph::buffer::list cbl; |
11fdf7f2 TL |
3451 | decode(cbl, bl); |
3452 | auto cblp = cbl.cbegin(); | |
7c673cae | 3453 | crush->decode(cblp); |
11fdf7f2 TL |
3454 | // added in firefly; version increased in luminous, so it affects |
3455 | // giant, hammer, infernallis, jewel, and kraken. probably should be left | |
3456 | // alone until we require clients to be all luminous? | |
7c673cae | 3457 | if (struct_v >= 3) { |
11fdf7f2 | 3458 | decode(erasure_code_profiles, bl); |
7c673cae FG |
3459 | } else { |
3460 | erasure_code_profiles.clear(); | |
3461 | } | |
11fdf7f2 TL |
3462 | // version increased from 3 to 4 still in luminous, so same as above |
3463 | // applies. | |
7c673cae | 3464 | if (struct_v >= 4) { |
11fdf7f2 TL |
3465 | decode(pg_upmap, bl); |
3466 | decode(pg_upmap_items, bl); | |
7c673cae FG |
3467 | } else { |
3468 | pg_upmap.clear(); | |
3469 | pg_upmap_items.clear(); | |
3470 | } | |
11fdf7f2 TL |
3471 | // again, version increased from 5 to 6 still in luminous, so above |
3472 | // applies. | |
31f18b77 | 3473 | if (struct_v >= 6) { |
11fdf7f2 TL |
3474 | decode(crush_version, bl); |
3475 | } | |
3476 | // version increase from 6 to 7 in mimic | |
3477 | if (struct_v >= 7) { | |
3478 | decode(new_removed_snaps, bl); | |
3479 | decode(new_purged_snaps, bl); | |
3480 | } | |
3481 | // version increase from 7 to 8, 8 to 9, in nautilus. | |
3482 | if (struct_v >= 9) { | |
3483 | decode(last_up_change, bl); | |
3484 | decode(last_in_change, bl); | |
31f18b77 | 3485 | } |
7c673cae FG |
3486 | DECODE_FINISH(bl); // client-usable data |
3487 | } | |
3488 | ||
3489 | { | |
f67539c2 | 3490 | DECODE_START(10, bl); // extended, osd-only data |
11fdf7f2 TL |
3491 | decode(osd_addrs->hb_back_addrs, bl); |
3492 | decode(osd_info, bl); | |
f67539c2 | 3493 | decode(blocklist, bl); |
11fdf7f2 TL |
3494 | decode(osd_addrs->cluster_addrs, bl); |
3495 | decode(cluster_snapshot_epoch, bl); | |
3496 | decode(cluster_snapshot, bl); | |
3497 | decode(*osd_uuid, bl); | |
3498 | decode(osd_xinfo, bl); | |
3499 | decode(osd_addrs->hb_front_addrs, bl); | |
3500 | // | |
7c673cae | 3501 | if (struct_v >= 2) { |
11fdf7f2 TL |
3502 | decode(nearfull_ratio, bl); |
3503 | decode(full_ratio, bl); | |
7c673cae FG |
3504 | } else { |
3505 | nearfull_ratio = 0; | |
3506 | full_ratio = 0; | |
3507 | } | |
3508 | if (struct_v >= 3) { | |
11fdf7f2 | 3509 | decode(backfillfull_ratio, bl); |
7c673cae FG |
3510 | } else { |
3511 | backfillfull_ratio = 0; | |
3512 | } | |
31f18b77 FG |
3513 | if (struct_v == 4) { |
3514 | string r; | |
11fdf7f2 | 3515 | decode(r, bl); |
31f18b77 FG |
3516 | if (r.length()) |
3517 | require_min_compat_client = ceph_release_from_name(r.c_str()); | |
3518 | } | |
3519 | if (struct_v >= 5) { | |
11fdf7f2 TL |
3520 | decode(require_min_compat_client, bl); |
3521 | decode(require_osd_release, bl); | |
9f95a23c | 3522 | if (require_osd_release >= ceph_release_t::nautilus) { |
11fdf7f2 TL |
3523 | flags |= CEPH_OSDMAP_PGLOG_HARDLIMIT; |
3524 | } | |
9f95a23c | 3525 | if (require_osd_release >= ceph_release_t::luminous) { |
31f18b77 | 3526 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3527 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 FG |
3528 | } |
3529 | } else { | |
3530 | if (flags & CEPH_OSDMAP_REQUIRE_LUMINOUS) { | |
3531 | // only for compat with post-kraken pre-luminous test clusters | |
9f95a23c | 3532 | require_osd_release = ceph_release_t::luminous; |
31f18b77 | 3533 | flags &= ~(CEPH_OSDMAP_LEGACY_REQUIRE_FLAGS); |
c07f9fc5 | 3534 | flags |= CEPH_OSDMAP_RECOVERY_DELETES; |
31f18b77 | 3535 | } else if (flags & CEPH_OSDMAP_REQUIRE_KRAKEN) { |
9f95a23c | 3536 | require_osd_release = ceph_release_t::kraken; |
31f18b77 | 3537 | } else if (flags & CEPH_OSDMAP_REQUIRE_JEWEL) { |
9f95a23c | 3538 | require_osd_release = ceph_release_t::jewel; |
31f18b77 | 3539 | } else { |
9f95a23c | 3540 | require_osd_release = ceph_release_t::unknown; |
31f18b77 FG |
3541 | } |
3542 | } | |
11fdf7f2 TL |
3543 | if (struct_v >= 6) { |
3544 | decode(removed_snaps_queue, bl); | |
3545 | } | |
81eedcae TL |
3546 | if (struct_v >= 8) { |
3547 | decode(crush_node_flags, bl); | |
3548 | } else { | |
3549 | crush_node_flags.clear(); | |
3550 | } | |
3551 | if (struct_v >= 9) { | |
3552 | decode(device_class_flags, bl); | |
3553 | } else { | |
3554 | device_class_flags.clear(); | |
3555 | } | |
f67539c2 TL |
3556 | if (struct_v >= 10) { |
3557 | decode(stretch_mode_enabled, bl); | |
3558 | decode(stretch_bucket_count, bl); | |
3559 | decode(degraded_stretch_mode, bl); | |
3560 | decode(recovering_stretch_mode, bl); | |
3561 | decode(stretch_mode_bucket, bl); | |
3562 | } else { | |
3563 | stretch_mode_enabled = false; | |
3564 | stretch_bucket_count = 0; | |
3565 | degraded_stretch_mode = 0; | |
3566 | recovering_stretch_mode = 0; | |
3567 | stretch_mode_bucket = 0; | |
3568 | } | |
33c7a0ef TL |
3569 | if (struct_v >= 11) { |
3570 | decode(range_blocklist, bl); | |
3571 | calculated_ranges.clear(); | |
3572 | for (const auto& i : range_blocklist) { | |
3573 | calculated_ranges.emplace(i.first, i.first); | |
3574 | } | |
3575 | } | |
7c673cae FG |
3576 | DECODE_FINISH(bl); // osd-only data |
3577 | } | |
3578 | ||
3579 | if (struct_v >= 8) { | |
3580 | crc_front.substr_of(bl.get_bl(), start_offset, bl.get_off() - start_offset); | |
11fdf7f2 | 3581 | decode(crc, bl); |
7c673cae FG |
3582 | tail_offset = bl.get_off(); |
3583 | crc_defined = true; | |
3584 | } else { | |
3585 | crc_defined = false; | |
3586 | crc = 0; | |
3587 | } | |
3588 | ||
3589 | DECODE_FINISH(bl); // wrapper | |
3590 | ||
3591 | if (tail_offset) { | |
3592 | // verify crc | |
3593 | uint32_t actual = crc_front.crc32c(-1); | |
3594 | if (tail_offset < bl.get_off()) { | |
9f95a23c | 3595 | ceph::buffer::list tail; |
7c673cae FG |
3596 | tail.substr_of(bl.get_bl(), tail_offset, bl.get_off() - tail_offset); |
3597 | actual = tail.crc32c(actual); | |
3598 | } | |
3599 | if (crc != actual) { | |
3600 | ostringstream ss; | |
3601 | ss << "bad crc, actual " << actual << " != expected " << crc; | |
3602 | string s = ss.str(); | |
9f95a23c | 3603 | throw ceph::buffer::malformed_input(s.c_str()); |
7c673cae FG |
3604 | } |
3605 | } | |
3606 | ||
3607 | post_decode(); | |
3608 | } | |
3609 | ||
3610 | void OSDMap::post_decode() | |
3611 | { | |
3612 | // index pool names | |
3613 | name_pool.clear(); | |
3614 | for (const auto &pname : pool_name) { | |
3615 | name_pool[pname.second] = pname.first; | |
3616 | } | |
3617 | ||
3618 | calc_num_osds(); | |
3619 | _calc_up_osd_features(); | |
3620 | } | |
3621 | ||
3622 | void OSDMap::dump_erasure_code_profiles( | |
3623 | const mempool::osdmap::map<string,map<string,string>>& profiles, | |
3624 | Formatter *f) | |
3625 | { | |
3626 | f->open_object_section("erasure_code_profiles"); | |
3627 | for (const auto &profile : profiles) { | |
3628 | f->open_object_section(profile.first.c_str()); | |
3629 | for (const auto &profm : profile.second) { | |
9f95a23c | 3630 | f->dump_string(profm.first.c_str(), profm.second); |
7c673cae FG |
3631 | } |
3632 | f->close_section(); | |
3633 | } | |
3634 | f->close_section(); | |
3635 | } | |
3636 | ||
9f95a23c TL |
3637 | void OSDMap::dump_osds(Formatter *f) const |
3638 | { | |
3639 | f->open_array_section("osds"); | |
3640 | for (int i=0; i<get_max_osd(); i++) { | |
3641 | if (exists(i)) { | |
3642 | dump_osd(i, f); | |
3643 | } | |
3644 | } | |
3645 | f->close_section(); | |
3646 | } | |
3647 | ||
3648 | void OSDMap::dump_osd(int id, Formatter *f) const | |
3649 | { | |
3650 | ceph_assert(f != nullptr); | |
3651 | if (!exists(id)) { | |
3652 | return; | |
3653 | } | |
3654 | ||
3655 | f->open_object_section("osd_info"); | |
3656 | f->dump_int("osd", id); | |
3657 | f->dump_stream("uuid") << get_uuid(id); | |
3658 | f->dump_int("up", is_up(id)); | |
3659 | f->dump_int("in", is_in(id)); | |
3660 | f->dump_float("weight", get_weightf(id)); | |
3661 | f->dump_float("primary_affinity", get_primary_affinityf(id)); | |
3662 | get_info(id).dump(f); | |
3663 | f->dump_object("public_addrs", get_addrs(id)); | |
3664 | f->dump_object("cluster_addrs", get_cluster_addrs(id)); | |
3665 | f->dump_object("heartbeat_back_addrs", get_hb_back_addrs(id)); | |
3666 | f->dump_object("heartbeat_front_addrs", get_hb_front_addrs(id)); | |
3667 | // compat | |
3668 | f->dump_stream("public_addr") << get_addrs(id).get_legacy_str(); | |
3669 | f->dump_stream("cluster_addr") << get_cluster_addrs(id).get_legacy_str(); | |
3670 | f->dump_stream("heartbeat_back_addr") | |
3671 | << get_hb_back_addrs(id).get_legacy_str(); | |
3672 | f->dump_stream("heartbeat_front_addr") | |
3673 | << get_hb_front_addrs(id).get_legacy_str(); | |
3674 | ||
3675 | set<string> st; | |
3676 | get_state(id, st); | |
3677 | f->open_array_section("state"); | |
3678 | for (const auto &state : st) | |
3679 | f->dump_string("state", state); | |
3680 | f->close_section(); | |
3681 | ||
3682 | f->close_section(); | |
3683 | } | |
3684 | ||
7c673cae FG |
3685 | void OSDMap::dump(Formatter *f) const |
3686 | { | |
3687 | f->dump_int("epoch", get_epoch()); | |
3688 | f->dump_stream("fsid") << get_fsid(); | |
3689 | f->dump_stream("created") << get_created(); | |
3690 | f->dump_stream("modified") << get_modified(); | |
11fdf7f2 TL |
3691 | f->dump_stream("last_up_change") << last_up_change; |
3692 | f->dump_stream("last_in_change") << last_in_change; | |
7c673cae | 3693 | f->dump_string("flags", get_flag_string()); |
11fdf7f2 TL |
3694 | f->dump_unsigned("flags_num", flags); |
3695 | f->open_array_section("flags_set"); | |
3696 | set<string> flagset; | |
3697 | get_flag_set(&flagset); | |
3698 | for (auto p : flagset) { | |
3699 | f->dump_string("flag", p); | |
3700 | } | |
3701 | f->close_section(); | |
31f18b77 | 3702 | f->dump_unsigned("crush_version", get_crush_version()); |
7c673cae FG |
3703 | f->dump_float("full_ratio", full_ratio); |
3704 | f->dump_float("backfillfull_ratio", backfillfull_ratio); | |
3705 | f->dump_float("nearfull_ratio", nearfull_ratio); | |
3706 | f->dump_string("cluster_snapshot", get_cluster_snapshot()); | |
3707 | f->dump_int("pool_max", get_pool_max()); | |
3708 | f->dump_int("max_osd", get_max_osd()); | |
31f18b77 | 3709 | f->dump_string("require_min_compat_client", |
f67539c2 | 3710 | to_string(require_min_compat_client)); |
31f18b77 | 3711 | f->dump_string("min_compat_client", |
f67539c2 | 3712 | to_string(get_min_compat_client())); |
31f18b77 | 3713 | f->dump_string("require_osd_release", |
f67539c2 | 3714 | to_string(require_osd_release)); |
7c673cae FG |
3715 | |
3716 | f->open_array_section("pools"); | |
3717 | for (const auto &pool : pools) { | |
3718 | std::string name("<unknown>"); | |
3719 | const auto &pni = pool_name.find(pool.first); | |
3720 | if (pni != pool_name.end()) | |
3721 | name = pni->second; | |
3722 | f->open_object_section("pool"); | |
3723 | f->dump_int("pool", pool.first); | |
3724 | f->dump_string("pool_name", name); | |
3725 | pool.second.dump(f); | |
3726 | f->close_section(); | |
3727 | } | |
3728 | f->close_section(); | |
3729 | ||
9f95a23c | 3730 | dump_osds(f); |
7c673cae FG |
3731 | |
3732 | f->open_array_section("osd_xinfo"); | |
3733 | for (int i=0; i<get_max_osd(); i++) { | |
3734 | if (exists(i)) { | |
3735 | f->open_object_section("xinfo"); | |
3736 | f->dump_int("osd", i); | |
3737 | osd_xinfo[i].dump(f); | |
3738 | f->close_section(); | |
3739 | } | |
3740 | } | |
3741 | f->close_section(); | |
3742 | ||
3743 | f->open_array_section("pg_upmap"); | |
3744 | for (auto& p : pg_upmap) { | |
3745 | f->open_object_section("mapping"); | |
3746 | f->dump_stream("pgid") << p.first; | |
3747 | f->open_array_section("osds"); | |
3748 | for (auto q : p.second) { | |
3749 | f->dump_int("osd", q); | |
3750 | } | |
3751 | f->close_section(); | |
3752 | f->close_section(); | |
3753 | } | |
3754 | f->close_section(); | |
3755 | f->open_array_section("pg_upmap_items"); | |
3756 | for (auto& p : pg_upmap_items) { | |
3757 | f->open_object_section("mapping"); | |
3758 | f->dump_stream("pgid") << p.first; | |
3759 | f->open_array_section("mappings"); | |
3760 | for (auto& q : p.second) { | |
3761 | f->open_object_section("mapping"); | |
3762 | f->dump_int("from", q.first); | |
3763 | f->dump_int("to", q.second); | |
3764 | f->close_section(); | |
3765 | } | |
3766 | f->close_section(); | |
3767 | f->close_section(); | |
3768 | } | |
3769 | f->close_section(); | |
3770 | f->open_array_section("pg_temp"); | |
31f18b77 | 3771 | pg_temp->dump(f); |
7c673cae FG |
3772 | f->close_section(); |
3773 | ||
3774 | f->open_array_section("primary_temp"); | |
3775 | for (const auto &pg : *primary_temp) { | |
3776 | f->dump_stream("pgid") << pg.first; | |
3777 | f->dump_int("osd", pg.second); | |
3778 | } | |
3779 | f->close_section(); // primary_temp | |
3780 | ||
f67539c2 TL |
3781 | f->open_object_section("blocklist"); |
3782 | for (const auto &addr : blocklist) { | |
7c673cae FG |
3783 | stringstream ss; |
3784 | ss << addr.first; | |
3785 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3786 | } | |
3787 | f->close_section(); | |
33c7a0ef TL |
3788 | f->open_object_section("range_blocklist"); |
3789 | for (const auto &addr : range_blocklist) { | |
3790 | stringstream ss; | |
3791 | ss << addr.first; | |
3792 | f->dump_stream(ss.str().c_str()) << addr.second; | |
3793 | } | |
3794 | f->close_section(); | |
7c673cae FG |
3795 | |
3796 | dump_erasure_code_profiles(erasure_code_profiles, f); | |
11fdf7f2 TL |
3797 | |
3798 | f->open_array_section("removed_snaps_queue"); | |
3799 | for (auto& p : removed_snaps_queue) { | |
3800 | f->open_object_section("pool"); | |
3801 | f->dump_int("pool", p.first); | |
3802 | f->open_array_section("snaps"); | |
3803 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3804 | f->open_object_section("interval"); | |
3805 | f->dump_unsigned("begin", q.get_start()); | |
3806 | f->dump_unsigned("length", q.get_len()); | |
3807 | f->close_section(); | |
3808 | } | |
3809 | f->close_section(); | |
3810 | f->close_section(); | |
3811 | } | |
3812 | f->close_section(); | |
3813 | f->open_array_section("new_removed_snaps"); | |
3814 | for (auto& p : new_removed_snaps) { | |
3815 | f->open_object_section("pool"); | |
3816 | f->dump_int("pool", p.first); | |
3817 | f->open_array_section("snaps"); | |
3818 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3819 | f->open_object_section("interval"); | |
3820 | f->dump_unsigned("begin", q.get_start()); | |
3821 | f->dump_unsigned("length", q.get_len()); | |
3822 | f->close_section(); | |
3823 | } | |
3824 | f->close_section(); | |
3825 | f->close_section(); | |
3826 | } | |
3827 | f->close_section(); | |
3828 | f->open_array_section("new_purged_snaps"); | |
3829 | for (auto& p : new_purged_snaps) { | |
3830 | f->open_object_section("pool"); | |
3831 | f->dump_int("pool", p.first); | |
3832 | f->open_array_section("snaps"); | |
3833 | for (auto q = p.second.begin(); q != p.second.end(); ++q) { | |
3834 | f->open_object_section("interval"); | |
3835 | f->dump_unsigned("begin", q.get_start()); | |
3836 | f->dump_unsigned("length", q.get_len()); | |
3837 | f->close_section(); | |
3838 | } | |
3839 | f->close_section(); | |
3840 | f->close_section(); | |
3841 | } | |
3842 | f->close_section(); | |
81eedcae TL |
3843 | f->open_object_section("crush_node_flags"); |
3844 | for (auto& i : crush_node_flags) { | |
3845 | string s = crush->item_exists(i.first) ? crush->get_item_name(i.first) | |
3846 | : stringify(i.first); | |
3847 | f->open_array_section(s.c_str()); | |
3848 | set<string> st; | |
3849 | calc_state_set(i.second, st); | |
3850 | for (auto& j : st) { | |
3851 | f->dump_string("flag", j); | |
3852 | } | |
3853 | f->close_section(); | |
3854 | } | |
3855 | f->close_section(); | |
3856 | f->open_object_section("device_class_flags"); | |
3857 | for (auto& i : device_class_flags) { | |
3858 | const char* class_name = crush->get_class_name(i.first); | |
3859 | string s = class_name ? class_name : stringify(i.first); | |
3860 | f->open_array_section(s.c_str()); | |
3861 | set<string> st; | |
3862 | calc_state_set(i.second, st); | |
3863 | for (auto& j : st) { | |
3864 | f->dump_string("flag", j); | |
3865 | } | |
3866 | f->close_section(); | |
3867 | } | |
3868 | f->close_section(); | |
f67539c2 TL |
3869 | f->open_object_section("stretch_mode"); |
3870 | { | |
3871 | f->dump_bool("stretch_mode_enabled", stretch_mode_enabled); | |
3872 | f->dump_unsigned("stretch_bucket_count", stretch_bucket_count); | |
3873 | f->dump_unsigned("degraded_stretch_mode", degraded_stretch_mode); | |
3874 | f->dump_unsigned("recovering_stretch_mode", recovering_stretch_mode); | |
3875 | f->dump_int("stretch_mode_bucket", stretch_mode_bucket); | |
3876 | } | |
3877 | f->close_section(); | |
7c673cae FG |
3878 | } |
3879 | ||
3880 | void OSDMap::generate_test_instances(list<OSDMap*>& o) | |
3881 | { | |
3882 | o.push_back(new OSDMap); | |
3883 | ||
3884 | CephContext *cct = new CephContext(CODE_ENVIRONMENT_UTILITY); | |
3885 | o.push_back(new OSDMap); | |
3886 | uuid_d fsid; | |
224ce89b | 3887 | o.back()->build_simple(cct, 1, fsid, 16); |
7c673cae | 3888 | o.back()->created = o.back()->modified = utime_t(1, 2); // fix timestamp |
f67539c2 | 3889 | o.back()->blocklist[entity_addr_t()] = utime_t(5, 6); |
7c673cae FG |
3890 | cct->put(); |
3891 | } | |
3892 | ||
3893 | string OSDMap::get_flag_string(unsigned f) | |
3894 | { | |
3895 | string s; | |
7c673cae FG |
3896 | if (f & CEPH_OSDMAP_PAUSERD) |
3897 | s += ",pauserd"; | |
3898 | if (f & CEPH_OSDMAP_PAUSEWR) | |
3899 | s += ",pausewr"; | |
3900 | if (f & CEPH_OSDMAP_PAUSEREC) | |
3901 | s += ",pauserec"; | |
3902 | if (f & CEPH_OSDMAP_NOUP) | |
3903 | s += ",noup"; | |
3904 | if (f & CEPH_OSDMAP_NODOWN) | |
3905 | s += ",nodown"; | |
3906 | if (f & CEPH_OSDMAP_NOOUT) | |
3907 | s += ",noout"; | |
3908 | if (f & CEPH_OSDMAP_NOIN) | |
3909 | s += ",noin"; | |
3910 | if (f & CEPH_OSDMAP_NOBACKFILL) | |
3911 | s += ",nobackfill"; | |
3912 | if (f & CEPH_OSDMAP_NOREBALANCE) | |
3913 | s += ",norebalance"; | |
3914 | if (f & CEPH_OSDMAP_NORECOVER) | |
3915 | s += ",norecover"; | |
3916 | if (f & CEPH_OSDMAP_NOSCRUB) | |
3917 | s += ",noscrub"; | |
3918 | if (f & CEPH_OSDMAP_NODEEP_SCRUB) | |
3919 | s += ",nodeep-scrub"; | |
3920 | if (f & CEPH_OSDMAP_NOTIERAGENT) | |
3921 | s += ",notieragent"; | |
11fdf7f2 TL |
3922 | if (f & CEPH_OSDMAP_NOSNAPTRIM) |
3923 | s += ",nosnaptrim"; | |
7c673cae FG |
3924 | if (f & CEPH_OSDMAP_SORTBITWISE) |
3925 | s += ",sortbitwise"; | |
3926 | if (f & CEPH_OSDMAP_REQUIRE_JEWEL) | |
3927 | s += ",require_jewel_osds"; | |
3928 | if (f & CEPH_OSDMAP_REQUIRE_KRAKEN) | |
3929 | s += ",require_kraken_osds"; | |
3930 | if (f & CEPH_OSDMAP_REQUIRE_LUMINOUS) | |
3931 | s += ",require_luminous_osds"; | |
c07f9fc5 FG |
3932 | if (f & CEPH_OSDMAP_RECOVERY_DELETES) |
3933 | s += ",recovery_deletes"; | |
181888fb FG |
3934 | if (f & CEPH_OSDMAP_PURGED_SNAPDIRS) |
3935 | s += ",purged_snapdirs"; | |
f64942e4 AA |
3936 | if (f & CEPH_OSDMAP_PGLOG_HARDLIMIT) |
3937 | s += ",pglog_hardlimit"; | |
7c673cae FG |
3938 | if (s.length()) |
3939 | s.erase(0, 1); | |
3940 | return s; | |
3941 | } | |
3942 | ||
3943 | string OSDMap::get_flag_string() const | |
3944 | { | |
3945 | return get_flag_string(flags); | |
3946 | } | |
3947 | ||
7c673cae FG |
3948 | void OSDMap::print_pools(ostream& out) const |
3949 | { | |
3950 | for (const auto &pool : pools) { | |
3951 | std::string name("<unknown>"); | |
3952 | const auto &pni = pool_name.find(pool.first); | |
3953 | if (pni != pool_name.end()) | |
3954 | name = pni->second; | |
3955 | out << "pool " << pool.first | |
3956 | << " '" << name | |
3957 | << "' " << pool.second << "\n"; | |
3958 | ||
3959 | for (const auto &snap : pool.second.snaps) | |
3960 | out << "\tsnap " << snap.second.snapid << " '" << snap.second.name << "' " << snap.second.stamp << "\n"; | |
3961 | ||
3962 | if (!pool.second.removed_snaps.empty()) | |
3963 | out << "\tremoved_snaps " << pool.second.removed_snaps << "\n"; | |
11fdf7f2 TL |
3964 | auto p = removed_snaps_queue.find(pool.first); |
3965 | if (p != removed_snaps_queue.end()) { | |
3966 | out << "\tremoved_snaps_queue " << p->second << "\n"; | |
3967 | } | |
7c673cae FG |
3968 | } |
3969 | out << std::endl; | |
3970 | } | |
3971 | ||
9f95a23c TL |
3972 | void OSDMap::print_osds(ostream& out) const |
3973 | { | |
3974 | for (int i=0; i<get_max_osd(); i++) { | |
3975 | if (exists(i)) { | |
3976 | print_osd(i, out); | |
3977 | } | |
3978 | } | |
3979 | } | |
3980 | void OSDMap::print_osd(int id, ostream& out) const | |
3981 | { | |
3982 | if (!exists(id)) { | |
3983 | return; | |
3984 | } | |
3985 | ||
3986 | out << "osd." << id; | |
3987 | out << (is_up(id) ? " up ":" down"); | |
3988 | out << (is_in(id) ? " in ":" out"); | |
3989 | out << " weight " << get_weightf(id); | |
3990 | if (get_primary_affinity(id) != CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { | |
3991 | out << " primary_affinity " << get_primary_affinityf(id); | |
3992 | } | |
3993 | const osd_info_t& info(get_info(id)); | |
3994 | out << " " << info; | |
3995 | out << " " << get_addrs(id) << " " << get_cluster_addrs(id); | |
3996 | set<string> st; | |
3997 | get_state(id, st); | |
3998 | out << " " << st; | |
3999 | if (!get_uuid(id).is_zero()) { | |
4000 | out << " " << get_uuid(id); | |
4001 | } | |
4002 | out << "\n"; | |
4003 | } | |
4004 | ||
7c673cae FG |
4005 | void OSDMap::print(ostream& out) const |
4006 | { | |
4007 | out << "epoch " << get_epoch() << "\n" | |
4008 | << "fsid " << get_fsid() << "\n" | |
4009 | << "created " << get_created() << "\n" | |
4010 | << "modified " << get_modified() << "\n"; | |
4011 | ||
4012 | out << "flags " << get_flag_string() << "\n"; | |
31f18b77 | 4013 | out << "crush_version " << get_crush_version() << "\n"; |
7c673cae FG |
4014 | out << "full_ratio " << full_ratio << "\n"; |
4015 | out << "backfillfull_ratio " << backfillfull_ratio << "\n"; | |
4016 | out << "nearfull_ratio " << nearfull_ratio << "\n"; | |
9f95a23c | 4017 | if (require_min_compat_client != ceph_release_t::unknown) { |
31f18b77 | 4018 | out << "require_min_compat_client " |
9f95a23c | 4019 | << require_min_compat_client << "\n"; |
7c673cae | 4020 | } |
9f95a23c | 4021 | out << "min_compat_client " << get_min_compat_client() |
31f18b77 | 4022 | << "\n"; |
9f95a23c TL |
4023 | if (require_osd_release > ceph_release_t::unknown) { |
4024 | out << "require_osd_release " << require_osd_release | |
224ce89b WB |
4025 | << "\n"; |
4026 | } | |
f67539c2 TL |
4027 | out << "stretch_mode_enabled " << (stretch_mode_enabled ? "true" : "false") << "\n"; |
4028 | if (stretch_mode_enabled) { | |
4029 | out << "stretch_bucket_count " << stretch_bucket_count << "\n"; | |
4030 | out << "degraded_stretch_mode " << degraded_stretch_mode << "\n"; | |
4031 | out << "recovering_stretch_mode " << recovering_stretch_mode << "\n"; | |
4032 | out << "stretch_mode_bucket " << stretch_mode_bucket << "\n"; | |
4033 | } | |
7c673cae FG |
4034 | if (get_cluster_snapshot().length()) |
4035 | out << "cluster_snapshot " << get_cluster_snapshot() << "\n"; | |
4036 | out << "\n"; | |
4037 | ||
4038 | print_pools(out); | |
4039 | ||
4040 | out << "max_osd " << get_max_osd() << "\n"; | |
9f95a23c | 4041 | print_osds(out); |
7c673cae FG |
4042 | out << std::endl; |
4043 | ||
4044 | for (auto& p : pg_upmap) { | |
4045 | out << "pg_upmap " << p.first << " " << p.second << "\n"; | |
4046 | } | |
4047 | for (auto& p : pg_upmap_items) { | |
4048 | out << "pg_upmap_items " << p.first << " " << p.second << "\n"; | |
4049 | } | |
4050 | ||
f67539c2 | 4051 | for (const auto& pg : *pg_temp) |
7c673cae FG |
4052 | out << "pg_temp " << pg.first << " " << pg.second << "\n"; |
4053 | ||
f67539c2 | 4054 | for (const auto& pg : *primary_temp) |
7c673cae FG |
4055 | out << "primary_temp " << pg.first << " " << pg.second << "\n"; |
4056 | ||
f67539c2 TL |
4057 | for (const auto &addr : blocklist) |
4058 | out << "blocklist " << addr.first << " expires " << addr.second << "\n"; | |
33c7a0ef TL |
4059 | for (const auto &addr : range_blocklist) |
4060 | out << "range blocklist " << addr.first << " expires " << addr.second << "\n"; | |
7c673cae FG |
4061 | } |
4062 | ||
4063 | class OSDTreePlainDumper : public CrushTreeDumper::Dumper<TextTable> { | |
4064 | public: | |
4065 | typedef CrushTreeDumper::Dumper<TextTable> Parent; | |
31f18b77 FG |
4066 | |
4067 | OSDTreePlainDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
4068 | unsigned f) | |
c07f9fc5 | 4069 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
4070 | |
4071 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
4072 | if (!filter) { |
4073 | return true; // normal case | |
4074 | } | |
4075 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
4076 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
4077 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
4078 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
4079 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
4080 | return true; | |
31f18b77 | 4081 | } |
c07f9fc5 | 4082 | return false; |
31f18b77 FG |
4083 | } |
4084 | ||
4085 | bool should_dump_empty_bucket() const override { | |
4086 | return !filter; | |
4087 | } | |
7c673cae | 4088 | |
11fdf7f2 | 4089 | void init_table(TextTable *tbl) { |
7c673cae | 4090 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 4091 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae FG |
4092 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
4093 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
c07f9fc5 | 4094 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
7c673cae | 4095 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); |
224ce89b | 4096 | tbl->define_column("PRI-AFF", TextTable::LEFT, TextTable::RIGHT); |
11fdf7f2 TL |
4097 | } |
4098 | void dump(TextTable *tbl, string& bucket) { | |
4099 | init_table(tbl); | |
7c673cae | 4100 | |
11fdf7f2 TL |
4101 | if (!bucket.empty()) { |
4102 | set_root(bucket); | |
4103 | Parent::dump(tbl); | |
4104 | } else { | |
4105 | Parent::dump(tbl); | |
4106 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4107 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) { | |
4108 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), tbl); | |
4109 | } | |
31f18b77 | 4110 | } |
7c673cae FG |
4111 | } |
4112 | } | |
4113 | ||
4114 | protected: | |
4115 | void dump_item(const CrushTreeDumper::Item &qi, TextTable *tbl) override { | |
224ce89b WB |
4116 | const char *c = crush->get_item_class(qi.id); |
4117 | if (!c) | |
4118 | c = ""; | |
7c673cae | 4119 | *tbl << qi.id |
224ce89b | 4120 | << c |
7c673cae FG |
4121 | << weightf_t(qi.weight); |
4122 | ||
4123 | ostringstream name; | |
4124 | for (int k = 0; k < qi.depth; k++) | |
4125 | name << " "; | |
4126 | if (qi.is_bucket()) { | |
4127 | name << crush->get_type_name(crush->get_bucket_type(qi.id)) << " " | |
4128 | << crush->get_item_name(qi.id); | |
4129 | } else { | |
4130 | name << "osd." << qi.id; | |
4131 | } | |
4132 | *tbl << name.str(); | |
4133 | ||
4134 | if (!qi.is_bucket()) { | |
4135 | if (!osdmap->exists(qi.id)) { | |
4136 | *tbl << "DNE" | |
4137 | << 0; | |
4138 | } else { | |
c07f9fc5 FG |
4139 | string s; |
4140 | if (osdmap->is_up(qi.id)) { | |
4141 | s = "up"; | |
4142 | } else if (osdmap->is_destroyed(qi.id)) { | |
4143 | s = "destroyed"; | |
4144 | } else { | |
4145 | s = "down"; | |
4146 | } | |
4147 | *tbl << s | |
7c673cae FG |
4148 | << weightf_t(osdmap->get_weightf(qi.id)) |
4149 | << weightf_t(osdmap->get_primary_affinityf(qi.id)); | |
4150 | } | |
4151 | } | |
4152 | *tbl << TextTable::endrow; | |
4153 | } | |
4154 | ||
4155 | private: | |
4156 | const OSDMap *osdmap; | |
31f18b77 | 4157 | const unsigned filter; |
7c673cae FG |
4158 | }; |
4159 | ||
4160 | class OSDTreeFormattingDumper : public CrushTreeDumper::FormattingDumper { | |
4161 | public: | |
4162 | typedef CrushTreeDumper::FormattingDumper Parent; | |
4163 | ||
31f18b77 FG |
4164 | OSDTreeFormattingDumper(const CrushWrapper *crush, const OSDMap *osdmap_, |
4165 | unsigned f) | |
c07f9fc5 | 4166 | : Parent(crush, osdmap_->get_pool_names()), osdmap(osdmap_), filter(f) { } |
31f18b77 FG |
4167 | |
4168 | bool should_dump_leaf(int i) const override { | |
c07f9fc5 FG |
4169 | if (!filter) { |
4170 | return true; // normal case | |
4171 | } | |
4172 | if (((filter & OSDMap::DUMP_UP) && osdmap->is_up(i)) || | |
4173 | ((filter & OSDMap::DUMP_DOWN) && osdmap->is_down(i)) || | |
4174 | ((filter & OSDMap::DUMP_IN) && osdmap->is_in(i)) || | |
4175 | ((filter & OSDMap::DUMP_OUT) && osdmap->is_out(i)) || | |
4176 | ((filter & OSDMap::DUMP_DESTROYED) && osdmap->is_destroyed(i))) { | |
4177 | return true; | |
31f18b77 | 4178 | } |
c07f9fc5 | 4179 | return false; |
31f18b77 FG |
4180 | } |
4181 | ||
4182 | bool should_dump_empty_bucket() const override { | |
4183 | return !filter; | |
4184 | } | |
7c673cae | 4185 | |
11fdf7f2 TL |
4186 | void dump(Formatter *f, string& bucket) { |
4187 | if (!bucket.empty()) { | |
4188 | set_root(bucket); | |
4189 | f->open_array_section("nodes"); | |
4190 | Parent::dump(f); | |
4191 | f->close_section(); | |
4192 | } else { | |
4193 | f->open_array_section("nodes"); | |
4194 | Parent::dump(f); | |
4195 | f->close_section(); | |
4196 | f->open_array_section("stray"); | |
4197 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
4198 | if (osdmap->exists(i) && !is_touched(i) && should_dump_leaf(i)) | |
4199 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); | |
4200 | } | |
4201 | f->close_section(); | |
7c673cae | 4202 | } |
7c673cae FG |
4203 | } |
4204 | ||
4205 | protected: | |
4206 | void dump_item_fields(const CrushTreeDumper::Item &qi, Formatter *f) override { | |
4207 | Parent::dump_item_fields(qi, f); | |
4208 | if (!qi.is_bucket()) | |
4209 | { | |
c07f9fc5 FG |
4210 | string s; |
4211 | if (osdmap->is_up(qi.id)) { | |
4212 | s = "up"; | |
4213 | } else if (osdmap->is_destroyed(qi.id)) { | |
4214 | s = "destroyed"; | |
4215 | } else { | |
4216 | s = "down"; | |
4217 | } | |
7c673cae | 4218 | f->dump_unsigned("exists", (int)osdmap->exists(qi.id)); |
c07f9fc5 | 4219 | f->dump_string("status", s); |
7c673cae FG |
4220 | f->dump_float("reweight", osdmap->get_weightf(qi.id)); |
4221 | f->dump_float("primary_affinity", osdmap->get_primary_affinityf(qi.id)); | |
4222 | } | |
4223 | } | |
4224 | ||
4225 | private: | |
4226 | const OSDMap *osdmap; | |
31f18b77 | 4227 | const unsigned filter; |
7c673cae FG |
4228 | }; |
4229 | ||
11fdf7f2 | 4230 | void OSDMap::print_tree(Formatter *f, ostream *out, unsigned filter, string bucket) const |
7c673cae | 4231 | { |
31f18b77 | 4232 | if (f) { |
11fdf7f2 | 4233 | OSDTreeFormattingDumper(crush.get(), this, filter).dump(f, bucket); |
31f18b77 | 4234 | } else { |
11fdf7f2 | 4235 | ceph_assert(out); |
7c673cae | 4236 | TextTable tbl; |
11fdf7f2 | 4237 | OSDTreePlainDumper(crush.get(), this, filter).dump(&tbl, bucket); |
7c673cae FG |
4238 | *out << tbl; |
4239 | } | |
4240 | } | |
4241 | ||
224ce89b | 4242 | void OSDMap::print_summary(Formatter *f, ostream& out, |
11fdf7f2 | 4243 | const string& prefix, bool extra) const |
7c673cae FG |
4244 | { |
4245 | if (f) { | |
7c673cae FG |
4246 | f->dump_int("epoch", get_epoch()); |
4247 | f->dump_int("num_osds", get_num_osds()); | |
4248 | f->dump_int("num_up_osds", get_num_up_osds()); | |
9f95a23c | 4249 | f->dump_int("osd_up_since", last_up_change.to_msec() / 1000); |
7c673cae | 4250 | f->dump_int("num_in_osds", get_num_in_osds()); |
9f95a23c | 4251 | f->dump_int("osd_in_since", last_in_change.to_msec() / 1000); |
7c673cae | 4252 | f->dump_unsigned("num_remapped_pgs", get_num_pg_temp()); |
7c673cae | 4253 | } else { |
11fdf7f2 | 4254 | utime_t now = ceph_clock_now(); |
31f18b77 | 4255 | out << get_num_osds() << " osds: " |
11fdf7f2 TL |
4256 | << get_num_up_osds() << " up"; |
4257 | if (last_up_change != utime_t()) { | |
4258 | out << " (since " << utimespan_str(now - last_up_change) << ")"; | |
4259 | } | |
4260 | out << ", " << get_num_in_osds() << " in"; | |
4261 | if (last_in_change != utime_t()) { | |
4262 | out << " (since " << utimespan_str(now - last_in_change) << ")"; | |
4263 | } | |
4264 | if (extra) | |
4265 | out << "; epoch: e" << get_epoch(); | |
7c673cae FG |
4266 | if (get_num_pg_temp()) |
4267 | out << "; " << get_num_pg_temp() << " remapped pgs"; | |
4268 | out << "\n"; | |
4269 | uint64_t important_flags = flags & ~CEPH_OSDMAP_SEMIHIDDEN_FLAGS; | |
4270 | if (important_flags) | |
224ce89b | 4271 | out << prefix << "flags " << get_flag_string(important_flags) << "\n"; |
7c673cae FG |
4272 | } |
4273 | } | |
4274 | ||
4275 | void OSDMap::print_oneline_summary(ostream& out) const | |
4276 | { | |
4277 | out << "e" << get_epoch() << ": " | |
31f18b77 | 4278 | << get_num_osds() << " total, " |
7c673cae FG |
4279 | << get_num_up_osds() << " up, " |
4280 | << get_num_in_osds() << " in"; | |
7c673cae FG |
4281 | } |
4282 | ||
3efd9988 | 4283 | bool OSDMap::crush_rule_in_use(int rule_id) const |
7c673cae FG |
4284 | { |
4285 | for (const auto &pool : pools) { | |
3efd9988 | 4286 | if (pool.second.crush_rule == rule_id) |
7c673cae FG |
4287 | return true; |
4288 | } | |
4289 | return false; | |
4290 | } | |
4291 | ||
3efd9988 FG |
4292 | int OSDMap::validate_crush_rules(CrushWrapper *newcrush, |
4293 | ostream *ss) const | |
4294 | { | |
4295 | for (auto& i : pools) { | |
4296 | auto& pool = i.second; | |
4297 | int ruleno = pool.get_crush_rule(); | |
4298 | if (!newcrush->rule_exists(ruleno)) { | |
4299 | *ss << "pool " << i.first << " references crush_rule " << ruleno | |
4300 | << " but it is not present"; | |
4301 | return -EINVAL; | |
4302 | } | |
20effc67 | 4303 | if (newcrush->get_rule_type(ruleno) != (int)pool.get_type()) { |
3efd9988 FG |
4304 | *ss << "pool " << i.first << " type does not match rule " << ruleno; |
4305 | return -EINVAL; | |
4306 | } | |
3efd9988 FG |
4307 | } |
4308 | return 0; | |
4309 | } | |
4310 | ||
224ce89b WB |
4311 | int OSDMap::build_simple_optioned(CephContext *cct, epoch_t e, uuid_d &fsid, |
4312 | int nosd, int pg_bits, int pgp_bits, | |
4313 | bool default_pool) | |
7c673cae | 4314 | { |
224ce89b WB |
4315 | ldout(cct, 10) << "build_simple on " << nosd |
4316 | << " osds" << dendl; | |
7c673cae FG |
4317 | epoch = e; |
4318 | set_fsid(fsid); | |
4319 | created = modified = ceph_clock_now(); | |
4320 | ||
4321 | if (nosd >= 0) { | |
4322 | set_max_osd(nosd); | |
4323 | } else { | |
4324 | // count osds | |
4325 | int maxosd = 0; | |
11fdf7f2 | 4326 | const auto& conf = cct->_conf; |
7c673cae | 4327 | vector<string> sections; |
11fdf7f2 | 4328 | conf.get_all_sections(sections); |
7c673cae FG |
4329 | |
4330 | for (auto §ion : sections) { | |
4331 | if (section.find("osd.") != 0) | |
4332 | continue; | |
4333 | ||
4334 | const char *begin = section.c_str() + 4; | |
4335 | char *end = (char*)begin; | |
4336 | int o = strtol(begin, &end, 10); | |
4337 | if (*end != '\0') | |
4338 | continue; | |
4339 | ||
4340 | if (o > cct->_conf->mon_max_osd) { | |
4341 | lderr(cct) << "[osd." << o << "] in config has id > mon_max_osd " << cct->_conf->mon_max_osd << dendl; | |
4342 | return -ERANGE; | |
4343 | } | |
4344 | ||
4345 | if (o > maxosd) | |
4346 | maxosd = o; | |
4347 | } | |
4348 | ||
4349 | set_max_osd(maxosd + 1); | |
4350 | } | |
4351 | ||
7c673cae FG |
4352 | |
4353 | stringstream ss; | |
4354 | int r; | |
4355 | if (nosd >= 0) | |
4356 | r = build_simple_crush_map(cct, *crush, nosd, &ss); | |
4357 | else | |
4358 | r = build_simple_crush_map_from_conf(cct, *crush, &ss); | |
11fdf7f2 | 4359 | ceph_assert(r == 0); |
7c673cae FG |
4360 | |
4361 | int poolbase = get_max_osd() ? get_max_osd() : 1; | |
4362 | ||
20effc67 | 4363 | const int default_replicated_rule = crush->get_osd_pool_default_crush_replicated_rule(cct); |
11fdf7f2 | 4364 | ceph_assert(default_replicated_rule >= 0); |
7c673cae | 4365 | |
224ce89b WB |
4366 | if (default_pool) { |
4367 | // pgp_num <= pg_num | |
4368 | if (pgp_bits > pg_bits) | |
4369 | pgp_bits = pg_bits; | |
4370 | ||
4371 | vector<string> pool_names; | |
4372 | pool_names.push_back("rbd"); | |
4373 | for (auto &plname : pool_names) { | |
4374 | int64_t pool = ++pool_max; | |
4375 | pools[pool].type = pg_pool_t::TYPE_REPLICATED; | |
4376 | pools[pool].flags = cct->_conf->osd_pool_default_flags; | |
4377 | if (cct->_conf->osd_pool_default_flag_hashpspool) | |
4378 | pools[pool].set_flag(pg_pool_t::FLAG_HASHPSPOOL); | |
4379 | if (cct->_conf->osd_pool_default_flag_nodelete) | |
4380 | pools[pool].set_flag(pg_pool_t::FLAG_NODELETE); | |
4381 | if (cct->_conf->osd_pool_default_flag_nopgchange) | |
4382 | pools[pool].set_flag(pg_pool_t::FLAG_NOPGCHANGE); | |
4383 | if (cct->_conf->osd_pool_default_flag_nosizechange) | |
4384 | pools[pool].set_flag(pg_pool_t::FLAG_NOSIZECHANGE); | |
20effc67 TL |
4385 | if (cct->_conf->osd_pool_default_flag_bulk) |
4386 | pools[pool].set_flag(pg_pool_t::FLAG_BULK); | |
11fdf7f2 TL |
4387 | pools[pool].size = cct->_conf.get_val<uint64_t>("osd_pool_default_size"); |
4388 | pools[pool].min_size = cct->_conf.get_osd_pool_default_min_size( | |
4389 | pools[pool].size); | |
224ce89b WB |
4390 | pools[pool].crush_rule = default_replicated_rule; |
4391 | pools[pool].object_hash = CEPH_STR_HASH_RJENKINS; | |
4392 | pools[pool].set_pg_num(poolbase << pg_bits); | |
4393 | pools[pool].set_pgp_num(poolbase << pgp_bits); | |
11fdf7f2 TL |
4394 | pools[pool].set_pg_num_target(poolbase << pg_bits); |
4395 | pools[pool].set_pgp_num_target(poolbase << pgp_bits); | |
224ce89b | 4396 | pools[pool].last_change = epoch; |
c07f9fc5 FG |
4397 | pools[pool].application_metadata.insert( |
4398 | {pg_pool_t::APPLICATION_NAME_RBD, {}}); | |
9f95a23c TL |
4399 | if (auto m = pg_pool_t::get_pg_autoscale_mode_by_name( |
4400 | cct->_conf.get_val<string>("osd_pool_default_pg_autoscale_mode")); | |
4401 | m != pg_pool_t::pg_autoscale_mode_t::UNKNOWN) { | |
4402 | pools[pool].pg_autoscale_mode = m; | |
4403 | } else { | |
4404 | pools[pool].pg_autoscale_mode = pg_pool_t::pg_autoscale_mode_t::OFF; | |
4405 | } | |
224ce89b WB |
4406 | pool_name[pool] = plname; |
4407 | name_pool[plname] = pool; | |
4408 | } | |
7c673cae FG |
4409 | } |
4410 | ||
7c673cae FG |
4411 | map<string,string> profile_map; |
4412 | r = get_erasure_code_profile_default(cct, profile_map, &ss); | |
4413 | if (r < 0) { | |
4414 | lderr(cct) << ss.str() << dendl; | |
4415 | return r; | |
4416 | } | |
4417 | set_erasure_code_profile("default", profile_map); | |
4418 | return 0; | |
4419 | } | |
4420 | ||
4421 | int OSDMap::get_erasure_code_profile_default(CephContext *cct, | |
4422 | map<string,string> &profile_map, | |
4423 | ostream *ss) | |
4424 | { | |
11fdf7f2 | 4425 | int r = get_json_str_map(cct->_conf.get_val<string>("osd_pool_default_erasure_code_profile"), |
7c673cae FG |
4426 | *ss, |
4427 | &profile_map); | |
4428 | return r; | |
4429 | } | |
4430 | ||
4431 | int OSDMap::_build_crush_types(CrushWrapper& crush) | |
4432 | { | |
4433 | crush.set_type_name(0, "osd"); | |
4434 | crush.set_type_name(1, "host"); | |
4435 | crush.set_type_name(2, "chassis"); | |
4436 | crush.set_type_name(3, "rack"); | |
4437 | crush.set_type_name(4, "row"); | |
4438 | crush.set_type_name(5, "pdu"); | |
4439 | crush.set_type_name(6, "pod"); | |
4440 | crush.set_type_name(7, "room"); | |
4441 | crush.set_type_name(8, "datacenter"); | |
11fdf7f2 TL |
4442 | crush.set_type_name(9, "zone"); |
4443 | crush.set_type_name(10, "region"); | |
4444 | crush.set_type_name(11, "root"); | |
4445 | return 11; | |
7c673cae FG |
4446 | } |
4447 | ||
4448 | int OSDMap::build_simple_crush_map(CephContext *cct, CrushWrapper& crush, | |
4449 | int nosd, ostream *ss) | |
4450 | { | |
4451 | crush.create(); | |
4452 | ||
4453 | // root | |
4454 | int root_type = _build_crush_types(crush); | |
4455 | int rootid; | |
4456 | int r = crush.add_bucket(0, 0, CRUSH_HASH_DEFAULT, | |
4457 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4458 | ceph_assert(r == 0); |
7c673cae FG |
4459 | crush.set_item_name(rootid, "default"); |
4460 | ||
f67539c2 TL |
4461 | map<string,string> loc{ |
4462 | {"host", "localhost"}, | |
4463 | {"rack", "localrack"}, | |
4464 | {"root", "default"} | |
4465 | }; | |
7c673cae | 4466 | for (int o=0; o<nosd; o++) { |
7c673cae FG |
4467 | ldout(cct, 10) << " adding osd." << o << " at " << loc << dendl; |
4468 | char name[32]; | |
4469 | snprintf(name, sizeof(name), "osd.%d", o); | |
4470 | crush.insert_item(cct, o, 1.0, name, loc); | |
4471 | } | |
4472 | ||
31f18b77 | 4473 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4474 | |
4475 | crush.finalize(); | |
4476 | ||
4477 | return 0; | |
4478 | } | |
4479 | ||
4480 | int OSDMap::build_simple_crush_map_from_conf(CephContext *cct, | |
4481 | CrushWrapper& crush, | |
4482 | ostream *ss) | |
4483 | { | |
11fdf7f2 | 4484 | const auto& conf = cct->_conf; |
7c673cae FG |
4485 | |
4486 | crush.create(); | |
4487 | ||
4488 | // root | |
4489 | int root_type = _build_crush_types(crush); | |
4490 | int rootid; | |
4491 | int r = crush.add_bucket(0, 0, | |
4492 | CRUSH_HASH_DEFAULT, | |
4493 | root_type, 0, NULL, NULL, &rootid); | |
11fdf7f2 | 4494 | ceph_assert(r == 0); |
7c673cae FG |
4495 | crush.set_item_name(rootid, "default"); |
4496 | ||
4497 | // add osds | |
4498 | vector<string> sections; | |
11fdf7f2 | 4499 | conf.get_all_sections(sections); |
7c673cae FG |
4500 | |
4501 | for (auto §ion : sections) { | |
4502 | if (section.find("osd.") != 0) | |
4503 | continue; | |
4504 | ||
4505 | const char *begin = section.c_str() + 4; | |
4506 | char *end = (char*)begin; | |
4507 | int o = strtol(begin, &end, 10); | |
4508 | if (*end != '\0') | |
4509 | continue; | |
4510 | ||
4511 | string host, rack, row, room, dc, pool; | |
4512 | vector<string> sectiontmp; | |
4513 | sectiontmp.push_back("osd"); | |
4514 | sectiontmp.push_back(section); | |
11fdf7f2 TL |
4515 | conf.get_val_from_conf_file(sectiontmp, "host", host, false); |
4516 | conf.get_val_from_conf_file(sectiontmp, "rack", rack, false); | |
4517 | conf.get_val_from_conf_file(sectiontmp, "row", row, false); | |
4518 | conf.get_val_from_conf_file(sectiontmp, "room", room, false); | |
4519 | conf.get_val_from_conf_file(sectiontmp, "datacenter", dc, false); | |
4520 | conf.get_val_from_conf_file(sectiontmp, "root", pool, false); | |
7c673cae FG |
4521 | |
4522 | if (host.length() == 0) | |
4523 | host = "unknownhost"; | |
4524 | if (rack.length() == 0) | |
4525 | rack = "unknownrack"; | |
4526 | ||
4527 | map<string,string> loc; | |
4528 | loc["host"] = host; | |
4529 | loc["rack"] = rack; | |
4530 | if (row.size()) | |
4531 | loc["row"] = row; | |
4532 | if (room.size()) | |
4533 | loc["room"] = room; | |
4534 | if (dc.size()) | |
4535 | loc["datacenter"] = dc; | |
4536 | loc["root"] = "default"; | |
4537 | ||
4538 | ldout(cct, 5) << " adding osd." << o << " at " << loc << dendl; | |
4539 | crush.insert_item(cct, o, 1.0, section, loc); | |
4540 | } | |
4541 | ||
31f18b77 | 4542 | build_simple_crush_rules(cct, crush, "default", ss); |
7c673cae FG |
4543 | |
4544 | crush.finalize(); | |
4545 | ||
4546 | return 0; | |
4547 | } | |
4548 | ||
4549 | ||
31f18b77 FG |
4550 | int OSDMap::build_simple_crush_rules( |
4551 | CephContext *cct, | |
4552 | CrushWrapper& crush, | |
4553 | const string& root, | |
4554 | ostream *ss) | |
7c673cae | 4555 | { |
20effc67 | 4556 | int crush_rule = crush.get_osd_pool_default_crush_replicated_rule(cct); |
7c673cae FG |
4557 | string failure_domain = |
4558 | crush.get_type_name(cct->_conf->osd_crush_chooseleaf_type); | |
4559 | ||
7c673cae | 4560 | int r; |
31f18b77 | 4561 | r = crush.add_simple_rule_at( |
224ce89b | 4562 | "replicated_rule", root, failure_domain, "", |
31f18b77 FG |
4563 | "firstn", pg_pool_t::TYPE_REPLICATED, |
4564 | crush_rule, ss); | |
7c673cae FG |
4565 | if (r < 0) |
4566 | return r; | |
4567 | // do not add an erasure rule by default or else we will implicitly | |
4568 | // require the crush_v2 feature of clients | |
4569 | return 0; | |
4570 | } | |
4571 | ||
4572 | int OSDMap::summarize_mapping_stats( | |
4573 | OSDMap *newmap, | |
4574 | const set<int64_t> *pools, | |
4575 | std::string *out, | |
4576 | Formatter *f) const | |
4577 | { | |
4578 | set<int64_t> ls; | |
4579 | if (pools) { | |
4580 | ls = *pools; | |
4581 | } else { | |
4582 | for (auto &p : get_pools()) | |
4583 | ls.insert(p.first); | |
4584 | } | |
4585 | ||
4586 | unsigned total_pg = 0; | |
4587 | unsigned moved_pg = 0; | |
4588 | vector<unsigned> base_by_osd(get_max_osd(), 0); | |
4589 | vector<unsigned> new_by_osd(get_max_osd(), 0); | |
4590 | for (int64_t pool_id : ls) { | |
4591 | const pg_pool_t *pi = get_pg_pool(pool_id); | |
31f18b77 FG |
4592 | vector<int> up, up2; |
4593 | int up_primary; | |
7c673cae | 4594 | for (unsigned ps = 0; ps < pi->get_pg_num(); ++ps) { |
11fdf7f2 | 4595 | pg_t pgid(ps, pool_id); |
7c673cae | 4596 | total_pg += pi->get_size(); |
31f18b77 | 4597 | pg_to_up_acting_osds(pgid, &up, &up_primary, nullptr, nullptr); |
7c673cae FG |
4598 | for (int osd : up) { |
4599 | if (osd >= 0 && osd < get_max_osd()) | |
4600 | ++base_by_osd[osd]; | |
4601 | } | |
4602 | if (newmap) { | |
31f18b77 | 4603 | newmap->pg_to_up_acting_osds(pgid, &up2, &up_primary, nullptr, nullptr); |
7c673cae FG |
4604 | for (int osd : up2) { |
4605 | if (osd >= 0 && osd < get_max_osd()) | |
4606 | ++new_by_osd[osd]; | |
4607 | } | |
4608 | if (pi->type == pg_pool_t::TYPE_ERASURE) { | |
4609 | for (unsigned i=0; i<up.size(); ++i) { | |
4610 | if (up[i] != up2[i]) { | |
4611 | ++moved_pg; | |
4612 | } | |
4613 | } | |
4614 | } else if (pi->type == pg_pool_t::TYPE_REPLICATED) { | |
4615 | for (int osd : up) { | |
4616 | if (std::find(up2.begin(), up2.end(), osd) == up2.end()) { | |
4617 | ++moved_pg; | |
4618 | } | |
4619 | } | |
4620 | } else { | |
11fdf7f2 | 4621 | ceph_abort_msg("unhandled pool type"); |
7c673cae FG |
4622 | } |
4623 | } | |
4624 | } | |
4625 | } | |
4626 | ||
4627 | unsigned num_up_in = 0; | |
4628 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4629 | if (is_up(osd) && is_in(osd)) | |
4630 | ++num_up_in; | |
4631 | } | |
4632 | if (!num_up_in) { | |
4633 | return -EINVAL; | |
4634 | } | |
4635 | ||
4636 | float avg_pg = (float)total_pg / (float)num_up_in; | |
4637 | float base_stddev = 0, new_stddev = 0; | |
4638 | int min = -1, max = -1; | |
4639 | unsigned min_base_pg = 0, max_base_pg = 0; | |
4640 | unsigned min_new_pg = 0, max_new_pg = 0; | |
4641 | for (int osd = 0; osd < get_max_osd(); ++osd) { | |
4642 | if (is_up(osd) && is_in(osd)) { | |
4643 | float base_diff = (float)base_by_osd[osd] - avg_pg; | |
4644 | base_stddev += base_diff * base_diff; | |
4645 | float new_diff = (float)new_by_osd[osd] - avg_pg; | |
4646 | new_stddev += new_diff * new_diff; | |
4647 | if (min < 0 || base_by_osd[osd] < min_base_pg) { | |
4648 | min = osd; | |
4649 | min_base_pg = base_by_osd[osd]; | |
4650 | min_new_pg = new_by_osd[osd]; | |
4651 | } | |
4652 | if (max < 0 || base_by_osd[osd] > max_base_pg) { | |
4653 | max = osd; | |
4654 | max_base_pg = base_by_osd[osd]; | |
4655 | max_new_pg = new_by_osd[osd]; | |
4656 | } | |
4657 | } | |
4658 | } | |
4659 | base_stddev = sqrt(base_stddev / num_up_in); | |
4660 | new_stddev = sqrt(new_stddev / num_up_in); | |
4661 | ||
4662 | float edev = sqrt(avg_pg * (1.0 - (1.0 / (double)num_up_in))); | |
4663 | ||
4664 | ostringstream ss; | |
4665 | if (f) | |
4666 | f->open_object_section("utilization"); | |
4667 | if (newmap) { | |
4668 | if (f) { | |
4669 | f->dump_unsigned("moved_pgs", moved_pg); | |
4670 | f->dump_unsigned("total_pgs", total_pg); | |
4671 | } else { | |
4672 | float percent = 0; | |
4673 | if (total_pg) | |
4674 | percent = (float)moved_pg * 100.0 / (float)total_pg; | |
4675 | ss << "moved " << moved_pg << " / " << total_pg | |
4676 | << " (" << percent << "%)\n"; | |
4677 | } | |
4678 | } | |
4679 | if (f) { | |
4680 | f->dump_float("avg_pgs", avg_pg); | |
4681 | f->dump_float("std_dev", base_stddev); | |
4682 | f->dump_float("expected_baseline_std_dev", edev); | |
4683 | if (newmap) | |
4684 | f->dump_float("new_std_dev", new_stddev); | |
4685 | } else { | |
4686 | ss << "avg " << avg_pg << "\n"; | |
4687 | ss << "stddev " << base_stddev; | |
4688 | if (newmap) | |
4689 | ss << " -> " << new_stddev; | |
4690 | ss << " (expected baseline " << edev << ")\n"; | |
4691 | } | |
4692 | if (min >= 0) { | |
4693 | if (f) { | |
4694 | f->dump_unsigned("min_osd", min); | |
4695 | f->dump_unsigned("min_osd_pgs", min_base_pg); | |
4696 | if (newmap) | |
4697 | f->dump_unsigned("new_min_osd_pgs", min_new_pg); | |
4698 | } else { | |
4699 | ss << "min osd." << min << " with " << min_base_pg; | |
4700 | if (newmap) | |
4701 | ss << " -> " << min_new_pg; | |
4702 | ss << " pgs (" << (float)min_base_pg / avg_pg; | |
4703 | if (newmap) | |
4704 | ss << " -> " << (float)min_new_pg / avg_pg; | |
4705 | ss << " * mean)\n"; | |
4706 | } | |
4707 | } | |
4708 | if (max >= 0) { | |
4709 | if (f) { | |
4710 | f->dump_unsigned("max_osd", max); | |
4711 | f->dump_unsigned("max_osd_pgs", max_base_pg); | |
4712 | if (newmap) | |
4713 | f->dump_unsigned("new_max_osd_pgs", max_new_pg); | |
4714 | } else { | |
4715 | ss << "max osd." << max << " with " << max_base_pg; | |
4716 | if (newmap) | |
4717 | ss << " -> " << max_new_pg; | |
4718 | ss << " pgs (" << (float)max_base_pg / avg_pg; | |
4719 | if (newmap) | |
4720 | ss << " -> " << (float)max_new_pg / avg_pg; | |
4721 | ss << " * mean)\n"; | |
4722 | } | |
4723 | } | |
4724 | if (f) | |
4725 | f->close_section(); | |
4726 | if (out) | |
4727 | *out = ss.str(); | |
4728 | return 0; | |
4729 | } | |
4730 | ||
7c673cae FG |
4731 | bool OSDMap::try_pg_upmap( |
4732 | CephContext *cct, | |
4733 | pg_t pg, ///< pg to potentially remap | |
4734 | const set<int>& overfull, ///< osds we'd want to evacuate | |
4735 | const vector<int>& underfull, ///< osds to move to, in order of preference | |
92f5a8d4 | 4736 | const vector<int>& more_underfull, ///< more osds only slightly underfull |
7c673cae FG |
4737 | vector<int> *orig, |
4738 | vector<int> *out) ///< resulting alternative mapping | |
4739 | { | |
4740 | const pg_pool_t *pool = get_pg_pool(pg.pool()); | |
4741 | if (!pool) | |
4742 | return false; | |
20effc67 | 4743 | int rule = pool->get_crush_rule(); |
7c673cae FG |
4744 | if (rule < 0) |
4745 | return false; | |
4746 | ||
7c673cae FG |
4747 | // make sure there is something there to remap |
4748 | bool any = false; | |
4749 | for (auto osd : *orig) { | |
4750 | if (overfull.count(osd)) { | |
4751 | any = true; | |
4752 | break; | |
4753 | } | |
4754 | } | |
4755 | if (!any) { | |
4756 | return false; | |
4757 | } | |
4758 | ||
4759 | int r = crush->try_remap_rule( | |
4760 | cct, | |
4761 | rule, | |
4762 | pool->get_size(), | |
4763 | overfull, underfull, | |
92f5a8d4 | 4764 | more_underfull, |
7c673cae FG |
4765 | *orig, |
4766 | out); | |
4767 | if (r < 0) | |
4768 | return false; | |
4769 | if (*out == *orig) | |
4770 | return false; | |
4771 | return true; | |
4772 | } | |
4773 | ||
4774 | int OSDMap::calc_pg_upmaps( | |
4775 | CephContext *cct, | |
92f5a8d4 | 4776 | uint32_t max_deviation, |
7c673cae | 4777 | int max, |
a8e16298 | 4778 | const set<int64_t>& only_pools, |
20effc67 TL |
4779 | OSDMap::Incremental *pending_inc, |
4780 | std::random_device::result_type *p_seed) | |
7c673cae | 4781 | { |
a8e16298 | 4782 | ldout(cct, 10) << __func__ << " pools " << only_pools << dendl; |
20effc67 | 4783 | OSDMap tmp_osd_map; |
92f5a8d4 TL |
4784 | // Can't be less than 1 pg |
4785 | if (max_deviation < 1) | |
4786 | max_deviation = 1; | |
20effc67 | 4787 | tmp_osd_map.deepish_copy_from(*this); |
7c673cae | 4788 | int num_changed = 0; |
a8e16298 TL |
4789 | map<int,set<pg_t>> pgs_by_osd; |
4790 | int total_pgs = 0; | |
4791 | float osd_weight_total = 0; | |
4792 | map<int,float> osd_weight; | |
a8e16298 | 4793 | |
20effc67 TL |
4794 | if (max <= 0) { |
4795 | lderr(cct) << __func__ << " abort due to max <= 0" << dendl; | |
4796 | return 0; | |
a8e16298 | 4797 | } |
20effc67 TL |
4798 | |
4799 | osd_weight_total = build_pool_pgs_info(cct, only_pools, tmp_osd_map, | |
4800 | total_pgs, pgs_by_osd, osd_weight); | |
a8e16298 TL |
4801 | if (osd_weight_total == 0) { |
4802 | lderr(cct) << __func__ << " abort due to osd_weight_total == 0" << dendl; | |
4803 | return 0; | |
4804 | } | |
20effc67 | 4805 | |
a8e16298 TL |
4806 | float pgs_per_weight = total_pgs / osd_weight_total; |
4807 | ldout(cct, 10) << " osd_weight_total " << osd_weight_total << dendl; | |
4808 | ldout(cct, 10) << " pgs_per_weight " << pgs_per_weight << dendl; | |
7c673cae | 4809 | |
a8e16298 TL |
4810 | float stddev = 0; |
4811 | map<int,float> osd_deviation; // osd, deviation(pgs) | |
4812 | multimap<float,int> deviation_osd; // deviation(pgs), osd | |
20effc67 TL |
4813 | float cur_max_deviation = calc_deviations(cct, pgs_by_osd, osd_weight, pgs_per_weight, |
4814 | osd_deviation, deviation_osd, stddev); | |
4815 | ||
92f5a8d4 TL |
4816 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
4817 | if (cur_max_deviation <= max_deviation) { | |
a8e16298 TL |
4818 | ldout(cct, 10) << __func__ << " distribution is almost perfect" |
4819 | << dendl; | |
4820 | return 0; | |
4821 | } | |
20effc67 | 4822 | |
a8e16298 TL |
4823 | bool skip_overfull = false; |
4824 | auto aggressive = | |
11fdf7f2 | 4825 | cct->_conf.get_val<bool>("osd_calc_pg_upmaps_aggressively"); |
a8e16298 | 4826 | auto local_fallback_retries = |
11fdf7f2 | 4827 | cct->_conf.get_val<uint64_t>("osd_calc_pg_upmaps_local_fallback_retries"); |
20effc67 | 4828 | |
a8e16298 | 4829 | while (max--) { |
92f5a8d4 | 4830 | ldout(cct, 30) << "Top of loop #" << max+1 << dendl; |
a8e16298 TL |
4831 | // build overfull and underfull |
4832 | set<int> overfull; | |
92f5a8d4 TL |
4833 | set<int> more_overfull; |
4834 | bool using_more_overfull = false; | |
a8e16298 | 4835 | vector<int> underfull; |
92f5a8d4 | 4836 | vector<int> more_underfull; |
20effc67 TL |
4837 | fill_overfull_underfull(cct, deviation_osd, max_deviation, |
4838 | overfull, more_overfull, | |
4839 | underfull, more_underfull); | |
7c673cae | 4840 | |
92f5a8d4 TL |
4841 | if (underfull.empty() && overfull.empty()) { |
4842 | ldout(cct, 20) << __func__ << " failed to build overfull and underfull" << dendl; | |
7c673cae | 4843 | break; |
a8e16298 | 4844 | } |
92f5a8d4 TL |
4845 | if (overfull.empty() && !underfull.empty()) { |
4846 | ldout(cct, 20) << __func__ << " Using more_overfull since we still have underfull" << dendl; | |
4847 | overfull = more_overfull; | |
4848 | using_more_overfull = true; | |
4849 | } | |
7c673cae | 4850 | |
a8e16298 TL |
4851 | ldout(cct, 10) << " overfull " << overfull |
4852 | << " underfull " << underfull | |
4853 | << dendl; | |
4854 | set<pg_t> to_skip; | |
4855 | uint64_t local_fallback_retried = 0; | |
4856 | ||
4857 | retry: | |
4858 | ||
4859 | set<pg_t> to_unmap; | |
4860 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>> to_upmap; | |
4861 | auto temp_pgs_by_osd = pgs_by_osd; | |
4862 | // always start with fullest, break if we find any changes to make | |
7c673cae | 4863 | for (auto p = deviation_osd.rbegin(); p != deviation_osd.rend(); ++p) { |
92f5a8d4 | 4864 | if (skip_overfull && !underfull.empty()) { |
a8e16298 TL |
4865 | ldout(cct, 10) << " skipping overfull " << dendl; |
4866 | break; // fall through to check underfull | |
4867 | } | |
7c673cae | 4868 | int osd = p->second; |
31f18b77 | 4869 | float deviation = p->first; |
9f95a23c TL |
4870 | if (deviation < 0) { |
4871 | ldout(cct, 10) << " hitting underfull osds now" | |
4872 | << " when trying to remap overfull osds" | |
4873 | << dendl; | |
4874 | break; | |
4875 | } | |
7c673cae | 4876 | float target = osd_weight[osd] * pgs_per_weight; |
92f5a8d4 TL |
4877 | ldout(cct, 10) << " Overfull search osd." << osd |
4878 | << " target " << target | |
4879 | << " deviation " << deviation | |
4880 | << dendl; | |
a8e16298 | 4881 | ceph_assert(target > 0); |
92f5a8d4 | 4882 | if (!using_more_overfull && deviation <= max_deviation) { |
7c673cae | 4883 | ldout(cct, 10) << " osd." << osd |
a8e16298 TL |
4884 | << " target " << target |
4885 | << " deviation " << deviation | |
92f5a8d4 | 4886 | << " < max deviation " << max_deviation |
a8e16298 | 4887 | << dendl; |
7c673cae FG |
4888 | break; |
4889 | } | |
7c673cae | 4890 | |
a8e16298 TL |
4891 | vector<pg_t> pgs; |
4892 | pgs.reserve(pgs_by_osd[osd].size()); | |
4893 | for (auto& pg : pgs_by_osd[osd]) { | |
4894 | if (to_skip.count(pg)) | |
4895 | continue; | |
4896 | pgs.push_back(pg); | |
4897 | } | |
4898 | if (aggressive) { | |
4899 | // shuffle PG list so they all get equal (in)attention | |
20effc67 | 4900 | std::shuffle(pgs.begin(), pgs.end(), get_random_engine(cct, p_seed)); |
a8e16298 | 4901 | } |
7c673cae | 4902 | // look for remaps we can un-remap |
20effc67 TL |
4903 | if (try_drop_remap_overfull(cct, pgs, tmp_osd_map, osd, |
4904 | temp_pgs_by_osd, to_unmap, to_upmap)) | |
4905 | goto test_change; | |
7c673cae | 4906 | |
a8e16298 | 4907 | // try upmap |
7c673cae | 4908 | for (auto pg : pgs) { |
20effc67 TL |
4909 | auto temp_it = tmp_osd_map.pg_upmap.find(pg); |
4910 | if (temp_it != tmp_osd_map.pg_upmap.end()) { | |
a8e16298 TL |
4911 | // leave pg_upmap alone |
4912 | // it must be specified by admin since balancer does not | |
4913 | // support pg_upmap yet | |
4914 | ldout(cct, 10) << " " << pg << " already has pg_upmap " | |
4915 | << temp_it->second << ", skipping" | |
4916 | << dendl; | |
7c673cae FG |
4917 | continue; |
4918 | } | |
20effc67 | 4919 | auto pg_pool_size = tmp_osd_map.get_pg_pool_size(pg); |
a8e16298 TL |
4920 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; |
4921 | set<int> existing; | |
20effc67 TL |
4922 | auto it = tmp_osd_map.pg_upmap_items.find(pg); |
4923 | if (it != tmp_osd_map.pg_upmap_items.end()) { | |
4924 | auto& um_items = it->second; | |
4925 | if (um_items.size() >= (size_t)pg_pool_size) { | |
4926 | ldout(cct, 10) << " " << pg << " already has full-size pg_upmap_items " | |
4927 | << um_items << ", skipping" | |
4928 | << dendl; | |
4929 | continue; | |
4930 | } else { | |
4931 | ldout(cct, 10) << " " << pg << " already has pg_upmap_items " | |
4932 | << um_items | |
4933 | << dendl; | |
4934 | new_upmap_items = um_items; | |
4935 | // build existing too (for dedup) | |
4936 | for (auto [um_from, um_to] : um_items) { | |
4937 | existing.insert(um_from); | |
4938 | existing.insert(um_to); | |
4939 | } | |
4940 | } | |
a8e16298 TL |
4941 | // fall through |
4942 | // to see if we can append more remapping pairs | |
20effc67 | 4943 | } |
a8e16298 | 4944 | ldout(cct, 10) << " trying " << pg << dendl; |
494da23a | 4945 | vector<int> raw, orig, out; |
20effc67 | 4946 | tmp_osd_map.pg_to_raw_upmap(pg, &raw, &orig); // including existing upmaps too |
92f5a8d4 | 4947 | if (!try_pg_upmap(cct, pg, overfull, underfull, more_underfull, &orig, &out)) { |
7c673cae FG |
4948 | continue; |
4949 | } | |
a8e16298 | 4950 | ldout(cct, 10) << " " << pg << " " << orig << " -> " << out << dendl; |
7c673cae FG |
4951 | if (orig.size() != out.size()) { |
4952 | continue; | |
4953 | } | |
a8e16298 | 4954 | ceph_assert(orig != out); |
20effc67 | 4955 | int pos = find_best_remap(cct, orig, out, existing, osd_deviation); |
92f5a8d4 | 4956 | if (pos != -1) { |
a8e16298 TL |
4957 | // append new remapping pairs slowly |
4958 | // This way we can make sure that each tiny change will | |
4959 | // definitely make distribution of PGs converging to | |
4960 | // the perfect status. | |
20effc67 TL |
4961 | add_remap_pair(cct, orig[pos], out[pos], pg, (size_t)pg_pool_size, |
4962 | osd, existing, temp_pgs_by_osd, | |
4963 | new_upmap_items, to_upmap); | |
a8e16298 | 4964 | goto test_change; |
7c673cae | 4965 | } |
a8e16298 TL |
4966 | } |
4967 | } | |
7c673cae | 4968 | |
a8e16298 TL |
4969 | ceph_assert(!(to_unmap.size() || to_upmap.size())); |
4970 | ldout(cct, 10) << " failed to find any changes for overfull osds" | |
4971 | << dendl; | |
20effc67 TL |
4972 | for (auto& [deviation, osd] : deviation_osd) { |
4973 | if (std::find(underfull.begin(), underfull.end(), osd) == | |
a8e16298 TL |
4974 | underfull.end()) |
4975 | break; | |
a8e16298 TL |
4976 | float target = osd_weight[osd] * pgs_per_weight; |
4977 | ceph_assert(target > 0); | |
92f5a8d4 TL |
4978 | if (fabsf(deviation) < max_deviation) { |
4979 | // respect max_deviation too | |
a8e16298 TL |
4980 | ldout(cct, 10) << " osd." << osd |
4981 | << " target " << target | |
4982 | << " deviation " << deviation | |
92f5a8d4 TL |
4983 | << " -> absolute " << fabsf(deviation) |
4984 | << " < max " << max_deviation | |
a8e16298 TL |
4985 | << dendl; |
4986 | break; | |
4987 | } | |
4988 | // look for remaps we can un-remap | |
20effc67 TL |
4989 | candidates_t candidates = build_candidates(cct, tmp_osd_map, to_skip, |
4990 | only_pools, aggressive, p_seed); | |
4991 | if (try_drop_remap_underfull(cct, candidates, osd, temp_pgs_by_osd, | |
4992 | to_unmap, to_upmap)) { | |
4993 | goto test_change; | |
a8e16298 | 4994 | } |
7c673cae | 4995 | } |
a8e16298 TL |
4996 | |
4997 | ceph_assert(!(to_unmap.size() || to_upmap.size())); | |
4998 | ldout(cct, 10) << " failed to find any changes for underfull osds" | |
4999 | << dendl; | |
5000 | if (!aggressive) { | |
5001 | ldout(cct, 10) << " break due to aggressive mode not enabled" << dendl; | |
5002 | break; | |
5003 | } else if (!skip_overfull) { | |
5004 | // safe to quit because below here we know | |
5005 | // we've done checking both overfull and underfull osds.. | |
5006 | ldout(cct, 10) << " break due to not being able to find any" | |
5007 | << " further optimizations" | |
5008 | << dendl; | |
7c673cae FG |
5009 | break; |
5010 | } | |
a8e16298 TL |
5011 | // restart with fullest and do exhaustive searching |
5012 | skip_overfull = false; | |
5013 | continue; | |
5014 | ||
5015 | test_change: | |
5016 | ||
5017 | // test change, apply if change is good | |
5018 | ceph_assert(to_unmap.size() || to_upmap.size()); | |
5019 | float new_stddev = 0; | |
5020 | map<int,float> temp_osd_deviation; | |
5021 | multimap<float,int> temp_deviation_osd; | |
20effc67 TL |
5022 | float cur_max_deviation = calc_deviations(cct, temp_pgs_by_osd, osd_weight, |
5023 | pgs_per_weight, temp_osd_deviation, | |
5024 | temp_deviation_osd, new_stddev); | |
a8e16298 TL |
5025 | ldout(cct, 10) << " stddev " << stddev << " -> " << new_stddev << dendl; |
5026 | if (new_stddev >= stddev) { | |
5027 | if (!aggressive) { | |
5028 | ldout(cct, 10) << " break because stddev is not decreasing" | |
5029 | << " and aggressive mode is not enabled" | |
5030 | << dendl; | |
5031 | break; | |
5032 | } | |
5033 | local_fallback_retried++; | |
5034 | if (local_fallback_retried >= local_fallback_retries) { | |
5035 | // does not make progress | |
5036 | // flip *skip_overfull* so both overfull and underfull | |
5037 | // get equal (in)attention | |
5038 | skip_overfull = !skip_overfull; | |
5039 | ldout(cct, 10) << " hit local_fallback_retries " | |
5040 | << local_fallback_retries | |
5041 | << dendl; | |
5042 | continue; | |
5043 | } | |
5044 | for (auto& i : to_unmap) | |
5045 | to_skip.insert(i); | |
5046 | for (auto& i : to_upmap) | |
5047 | to_skip.insert(i.first); | |
5048 | ldout(cct, 20) << " local_fallback_retried " << local_fallback_retried | |
5049 | << " to_skip " << to_skip | |
5050 | << dendl; | |
5051 | goto retry; | |
5052 | } | |
5053 | ||
5054 | // ready to go | |
5055 | ceph_assert(new_stddev < stddev); | |
5056 | stddev = new_stddev; | |
5057 | pgs_by_osd = temp_pgs_by_osd; | |
5058 | osd_deviation = temp_osd_deviation; | |
5059 | deviation_osd = temp_deviation_osd; | |
20effc67 TL |
5060 | |
5061 | num_changed += pack_upmap_results(cct, to_unmap, to_upmap, tmp_osd_map, pending_inc); | |
5062 | ||
92f5a8d4 TL |
5063 | ldout(cct, 20) << " stdev " << stddev << " max_deviation " << cur_max_deviation << dendl; |
5064 | if (cur_max_deviation <= max_deviation) { | |
5065 | ldout(cct, 10) << __func__ << " Optimization plan is almost perfect" | |
5066 | << dendl; | |
5067 | break; | |
5068 | } | |
7c673cae | 5069 | } |
a8e16298 | 5070 | ldout(cct, 10) << " num_changed = " << num_changed << dendl; |
7c673cae FG |
5071 | return num_changed; |
5072 | } | |
31f18b77 | 5073 | |
20effc67 TL |
5074 | float OSDMap::build_pool_pgs_info ( |
5075 | CephContext *cct, | |
5076 | const std::set<int64_t>& only_pools, ///< [optional] restrict to pool | |
5077 | const OSDMap& tmp_osd_map, | |
5078 | int& total_pgs, | |
5079 | map<int,set<pg_t>>& pgs_by_osd, | |
5080 | map<int,float>& osd_weight) | |
5081 | { | |
5082 | // | |
5083 | // This function builds some data structures that are used by calc_pg_upmaps. | |
5084 | // Specifically it builds pgs_by_osd and osd_weight maps, updates total_pgs | |
5085 | // and returns the osd_weight_total | |
5086 | // | |
5087 | float osd_weight_total = 0.0; | |
5088 | for (auto& [pid, pdata] : pools) { | |
5089 | if (!only_pools.empty() && !only_pools.count(pid)) | |
5090 | continue; | |
5091 | for (unsigned ps = 0; ps < pdata.get_pg_num(); ++ps) { | |
5092 | pg_t pg(ps, pid); | |
5093 | vector<int> up; | |
5094 | tmp_osd_map.pg_to_up_acting_osds(pg, &up, nullptr, nullptr, nullptr); | |
5095 | ldout(cct, 20) << __func__ << " " << pg << " up " << up << dendl; | |
5096 | for (auto osd : up) { | |
5097 | if (osd != CRUSH_ITEM_NONE) | |
5098 | pgs_by_osd[osd].insert(pg); | |
5099 | } | |
5100 | } | |
5101 | total_pgs += pdata.get_size() * pdata.get_pg_num(); | |
5102 | ||
5103 | map<int,float> pmap; | |
5104 | int ruleno = pdata.get_crush_rule(); | |
5105 | tmp_osd_map.crush->get_rule_weight_osd_map(ruleno, &pmap); | |
5106 | ldout(cct,20) << __func__ << " pool " << pid | |
5107 | << " ruleno " << ruleno | |
5108 | << " weight-map " << pmap | |
5109 | << dendl; | |
5110 | for (auto [oid, oweight] : pmap) { | |
5111 | auto adjusted_weight = tmp_osd_map.get_weightf(oid) * oweight; | |
5112 | if (adjusted_weight == 0) { | |
5113 | continue; | |
5114 | } | |
5115 | osd_weight[oid] += adjusted_weight; | |
5116 | osd_weight_total += adjusted_weight; | |
5117 | } | |
5118 | } | |
5119 | for (auto& [oid, oweight] : osd_weight) { | |
5120 | int pgs = 0; | |
5121 | auto p = pgs_by_osd.find(oid); | |
5122 | if (p != pgs_by_osd.end()) | |
5123 | pgs = p->second.size(); | |
5124 | else | |
5125 | pgs_by_osd.emplace(oid, set<pg_t>()); | |
5126 | ldout(cct, 20) << " osd." << oid << " weight " << oweight | |
5127 | << " pgs " << pgs << dendl; | |
5128 | } | |
5129 | return osd_weight_total; | |
5130 | ||
5131 | } // return total weight of all OSDs | |
5132 | ||
5133 | float OSDMap::calc_deviations ( | |
5134 | CephContext *cct, | |
5135 | const map<int,set<pg_t>>& pgs_by_osd, | |
5136 | const map<int,float>& osd_weight, | |
5137 | float pgs_per_weight, | |
5138 | map<int,float>& osd_deviation, | |
5139 | multimap<float,int>& deviation_osd, | |
5140 | float& stddev) // return current max deviation | |
5141 | { | |
5142 | // | |
5143 | // This function calculates the 2 maps osd_deviation and deviation_osd which | |
5144 | // hold the deviation between the current number of PGs which map to an OSD | |
5145 | // and the optimal number. Ot also calculates the stddev of the deviations and | |
5146 | // returns the current max deviation. | |
5147 | // NOTE - the calculation is not exactly stddev it is actually sttdev^2 but as | |
5148 | // long as it is monotonic with stddev (and it is), it is sufficient for | |
5149 | // the balancer code. | |
5150 | // | |
5151 | float cur_max_deviation = 0.0; | |
5152 | stddev = 0.0; | |
5153 | for (auto& [oid, opgs] : pgs_by_osd) { | |
5154 | // make sure osd is still there (belongs to this crush-tree) | |
5155 | ceph_assert(osd_weight.count(oid)); | |
5156 | float target = osd_weight.at(oid) * pgs_per_weight; | |
5157 | float deviation = (float)opgs.size() - target; | |
5158 | ldout(cct, 20) << " osd." << oid | |
5159 | << "\tpgs " << opgs.size() | |
5160 | << "\ttarget " << target | |
5161 | << "\tdeviation " << deviation | |
5162 | << dendl; | |
5163 | osd_deviation[oid] = deviation; | |
5164 | deviation_osd.insert(make_pair(deviation, oid)); | |
5165 | stddev += deviation * deviation; | |
5166 | if (fabsf(deviation) > cur_max_deviation) | |
5167 | cur_max_deviation = fabsf(deviation); | |
5168 | } | |
5169 | return cur_max_deviation; | |
5170 | } | |
5171 | ||
5172 | void OSDMap::fill_overfull_underfull ( | |
5173 | CephContext *cct, | |
5174 | const std::multimap<float,int>& deviation_osd, | |
5175 | int max_deviation, | |
5176 | std::set<int>& overfull, | |
5177 | std::set<int>& more_overfull, | |
5178 | std::vector<int>& underfull, | |
5179 | std::vector<int>& more_underfull) | |
5180 | { | |
5181 | // | |
5182 | // This function just fills the overfull and underfull data structures for the | |
5183 | // use of calc_pg_upmaps | |
5184 | // | |
5185 | for (auto i = deviation_osd.rbegin(); i != deviation_osd.rend(); i++) { | |
5186 | auto& odev = i->first; | |
5187 | auto& oid = i->second; | |
5188 | ldout(cct, 30) << " check " << odev << " <= " << max_deviation << dendl; | |
5189 | if (odev <= 0) | |
5190 | break; | |
5191 | if (odev > max_deviation) { | |
5192 | ldout(cct, 30) << " add overfull osd." << oid << dendl; | |
5193 | overfull.insert(oid); | |
5194 | } else { | |
5195 | more_overfull.insert(oid); | |
5196 | } | |
5197 | } | |
5198 | ||
5199 | for (auto i = deviation_osd.begin(); i != deviation_osd.end(); i++) { | |
5200 | auto& odev = i->first; | |
5201 | auto& oid = i->second; | |
5202 | ldout(cct, 30) << " check " << odev << " >= " << -(int)max_deviation << dendl; | |
5203 | if (odev >= 0) | |
5204 | break; | |
5205 | if (odev < -(int)max_deviation) { | |
5206 | ldout(cct, 30) << " add underfull osd." << oid << dendl; | |
5207 | underfull.push_back(oid); | |
5208 | } else { | |
5209 | more_underfull.push_back(oid); | |
5210 | } | |
5211 | } | |
5212 | } | |
5213 | ||
5214 | int OSDMap::pack_upmap_results( | |
5215 | CephContext *cct, | |
5216 | const std::set<pg_t>& to_unmap, | |
5217 | const std::map<pg_t, mempool::osdmap::vector<std::pair<int, int>>>& to_upmap, | |
5218 | OSDMap& tmp_osd_map, | |
5219 | OSDMap::Incremental *pending_inc) | |
5220 | { | |
5221 | // | |
5222 | // This function takes the input from the local variables to_unmap and to_upmap | |
5223 | // and updates tmp_osd_map (so that another iteration can run) and pending_inc | |
5224 | // (so that the results are visible outside calc_pg_upmaps) | |
5225 | // | |
5226 | int num_changed = 0; | |
5227 | for (auto& i : to_unmap) { | |
5228 | ldout(cct, 10) << " unmap pg " << i << dendl; | |
5229 | ceph_assert(tmp_osd_map.pg_upmap_items.count(i)); | |
5230 | tmp_osd_map.pg_upmap_items.erase(i); | |
5231 | pending_inc->old_pg_upmap_items.insert(i); | |
5232 | ++num_changed; | |
5233 | } | |
5234 | for (auto& [pg, um_items] : to_upmap) { | |
5235 | ldout(cct, 10) << " upmap pg " << pg | |
5236 | << " new pg_upmap_items " << um_items | |
5237 | << dendl; | |
5238 | tmp_osd_map.pg_upmap_items[pg] = um_items; | |
5239 | pending_inc->new_pg_upmap_items[pg] = um_items; | |
5240 | ++num_changed; | |
5241 | } | |
5242 | ||
5243 | return num_changed; | |
5244 | } | |
5245 | ||
5246 | std::default_random_engine OSDMap::get_random_engine( | |
5247 | CephContext *cct, | |
5248 | std::random_device::result_type *p_seed) | |
5249 | { | |
5250 | // | |
5251 | // This function creates a random_engine to be used for shuffling. | |
5252 | // When p_seed == nullptr it generates random engine with a seed from /dev/random | |
5253 | // when p_seed is not null, it uses (*p_seed + seed_set) as the seed and | |
5254 | // increments seed_set. This is used in order to craete regression test without | |
5255 | // random effect on the results. | |
5256 | // | |
5257 | static std::random_device::result_type seed_set = 0; | |
5258 | std::random_device::result_type seed; | |
5259 | if (p_seed == nullptr) { | |
5260 | std::random_device rd; | |
5261 | seed = rd(); | |
5262 | } | |
5263 | else { | |
5264 | seed = *p_seed + seed_set; | |
5265 | ldout(cct, 30) << " Starting random engine with seed " | |
5266 | << seed << dendl; | |
5267 | seed_set++; | |
5268 | } | |
5269 | return std::default_random_engine{seed}; | |
5270 | } | |
5271 | ||
5272 | bool OSDMap::try_drop_remap_overfull( | |
5273 | CephContext *cct, | |
5274 | const std::vector<pg_t>& pgs, | |
5275 | const OSDMap& tmp_osd_map, | |
5276 | int osd, | |
5277 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5278 | set<pg_t>& to_unmap, | |
5279 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5280 | { | |
5281 | // | |
5282 | // This function tries to drop existimg upmap items which map data to overfull | |
5283 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5284 | // if it found an item that can be dropped, false if not. | |
5285 | // | |
5286 | for (auto pg : pgs) { | |
5287 | auto p = tmp_osd_map.pg_upmap_items.find(pg); | |
5288 | if (p == tmp_osd_map.pg_upmap_items.end()) | |
5289 | continue; | |
5290 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5291 | auto& pg_upmap_items = p->second; | |
5292 | for (auto um_pair : pg_upmap_items) { | |
5293 | auto& um_from = um_pair.first; | |
5294 | auto& um_to = um_pair.second; | |
5295 | if (um_to == osd) { | |
5296 | ldout(cct, 10) << " will try dropping existing" | |
5297 | << " remapping pair " | |
5298 | << um_from << " -> " << um_to | |
5299 | << " which remapped " << pg | |
5300 | << " into overfull osd." << osd | |
5301 | << dendl; | |
5302 | temp_pgs_by_osd[um_to].erase(pg); | |
5303 | temp_pgs_by_osd[um_from].insert(pg); | |
5304 | } else { | |
5305 | new_upmap_items.push_back(um_pair); | |
5306 | } | |
5307 | } | |
5308 | if (new_upmap_items.empty()) { | |
5309 | // drop whole item | |
5310 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5311 | << " remapped " << pg << " into overfull osd." << osd | |
5312 | << ", will try cancelling it entirely" | |
5313 | << dendl; | |
5314 | to_unmap.insert(pg); | |
5315 | return true; | |
5316 | } else if (new_upmap_items.size() != pg_upmap_items.size()) { | |
5317 | // drop single remapping pair, updating | |
5318 | ceph_assert(new_upmap_items.size() < pg_upmap_items.size()); | |
5319 | ldout(cct, 10) << " existing pg_upmap_items " << pg_upmap_items | |
5320 | << " remapped " << pg << " into overfull osd." << osd | |
5321 | << ", new_pg_upmap_items now " << new_upmap_items | |
5322 | << dendl; | |
5323 | to_upmap[pg] = new_upmap_items; | |
5324 | return true; | |
5325 | } | |
5326 | } | |
5327 | return false; | |
5328 | } | |
5329 | ||
5330 | bool OSDMap::try_drop_remap_underfull( | |
5331 | CephContext *cct, | |
5332 | const candidates_t& candidates, | |
5333 | int osd, | |
5334 | map<int,std::set<pg_t>>& temp_pgs_by_osd, | |
5335 | set<pg_t>& to_unmap, | |
5336 | map<pg_t, mempool::osdmap::vector<std::pair<int32_t,int32_t>>>& to_upmap) | |
5337 | { | |
5338 | // | |
5339 | // This function tries to drop existimg upmap items which map data from underfull | |
5340 | // OSDs. It updates temp_pgs_by_osd, to_unmap and to_upmap and rerturns true | |
5341 | // if it found an item that can be dropped, false if not. | |
5342 | // | |
5343 | for (auto& [pg, um_pairs] : candidates) { | |
5344 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items; | |
5345 | for (auto& ump : um_pairs) { | |
5346 | auto& um_from = ump.first; | |
5347 | auto& um_to = ump.second; | |
5348 | if (um_from == osd) { | |
5349 | ldout(cct, 10) << " will try dropping existing" | |
5350 | << " remapping pair " | |
5351 | << um_from << " -> " << um_to | |
5352 | << " which remapped " << pg | |
5353 | << " out from underfull osd." << osd | |
5354 | << dendl; | |
5355 | temp_pgs_by_osd[um_to].erase(pg); | |
5356 | temp_pgs_by_osd[um_from].insert(pg); | |
5357 | } else { | |
5358 | new_upmap_items.push_back(ump); | |
5359 | } | |
5360 | } | |
5361 | if (new_upmap_items.empty()) { | |
5362 | // drop whole item | |
5363 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5364 | << " remapped " << pg | |
5365 | << " out from underfull osd." << osd | |
5366 | << ", will try cancelling it entirely" | |
5367 | << dendl; | |
5368 | to_unmap.insert(pg); | |
5369 | return true; | |
5370 | } else if (new_upmap_items.size() != um_pairs.size()) { | |
5371 | // drop single remapping pair, updating | |
5372 | ceph_assert(new_upmap_items.size() < um_pairs.size()); | |
5373 | ldout(cct, 10) << " existing pg_upmap_items " << um_pairs | |
5374 | << " remapped " << pg | |
5375 | << " out from underfull osd." << osd | |
5376 | << ", new_pg_upmap_items now " << new_upmap_items | |
5377 | << dendl; | |
5378 | to_upmap[pg] = new_upmap_items; | |
5379 | return true; | |
5380 | } | |
5381 | } | |
5382 | return false; | |
5383 | } | |
5384 | ||
5385 | void OSDMap::add_remap_pair( | |
5386 | CephContext *cct, | |
5387 | int orig, | |
5388 | int out, | |
5389 | pg_t pg, | |
5390 | size_t pg_pool_size, | |
5391 | int osd, | |
5392 | set<int>& existing, | |
5393 | map<int,set<pg_t>>& temp_pgs_by_osd, | |
5394 | mempool::osdmap::vector<pair<int32_t,int32_t>> new_upmap_items, | |
5395 | map<pg_t, mempool::osdmap::vector<pair<int32_t,int32_t>>>& to_upmap) | |
5396 | { | |
5397 | // | |
5398 | // add a single remap pair (in pg <pg> remap osd from <orig> to <out>) to all | |
5399 | // the relevant data structures | |
5400 | // | |
5401 | ldout(cct, 10) << " will try adding new remapping pair " | |
5402 | << orig << " -> " << out << " for " << pg | |
5403 | << (orig != osd ? " NOT selected osd" : "") | |
5404 | << dendl; | |
5405 | existing.insert(orig); | |
5406 | existing.insert(out); | |
5407 | temp_pgs_by_osd[orig].erase(pg); | |
5408 | temp_pgs_by_osd[out].insert(pg); | |
5409 | ceph_assert(new_upmap_items.size() < pg_pool_size); | |
5410 | new_upmap_items.push_back(make_pair(orig, out)); | |
5411 | // append new remapping pairs slowly | |
5412 | // This way we can make sure that each tiny change will | |
5413 | // definitely make distribution of PGs converging to | |
5414 | // the perfect status. | |
5415 | to_upmap[pg] = new_upmap_items; | |
5416 | ||
5417 | } | |
5418 | ||
5419 | int OSDMap::find_best_remap ( | |
5420 | CephContext *cct, | |
5421 | const vector<int>& orig, | |
5422 | const vector<int>& out, | |
5423 | const set<int>& existing, | |
5424 | const map<int,float> osd_deviation) | |
5425 | { | |
5426 | // | |
5427 | // Find the best remap from the suggestions in orig and out - the best remap | |
5428 | // is the one which maps from the OSD with the largest deviatoion (from the | |
5429 | // OSDs which are part of orig) | |
5430 | // | |
5431 | int best_pos = -1; | |
5432 | float max_dev = 0; | |
5433 | for (unsigned i = 0; i < out.size(); ++i) { | |
5434 | if (orig[i] == out[i]) | |
5435 | continue; // skip invalid remappings | |
5436 | if (existing.count(orig[i]) || existing.count(out[i])) | |
5437 | continue; // we want new remappings only! | |
5438 | if (osd_deviation.at(orig[i]) > max_dev) { | |
5439 | max_dev = osd_deviation.at(orig[i]); | |
5440 | best_pos = i; | |
5441 | ldout(cct, 30) << "Max osd." << orig[i] << " pos " << i << " dev " << osd_deviation.at(orig[i]) << dendl; | |
5442 | } | |
5443 | } | |
5444 | return best_pos; | |
5445 | } | |
5446 | ||
5447 | OSDMap::candidates_t OSDMap::build_candidates( | |
5448 | CephContext *cct, | |
5449 | const OSDMap& tmp_osd_map, | |
5450 | const set<pg_t> to_skip, | |
5451 | const set<int64_t>& only_pools, | |
5452 | bool aggressive, | |
5453 | std::random_device::result_type *p_seed) | |
5454 | { | |
5455 | // | |
5456 | // build the candidates data structure | |
5457 | // | |
5458 | candidates_t candidates; | |
5459 | candidates.reserve(tmp_osd_map.pg_upmap_items.size()); | |
5460 | for (auto& [pg, um_pair] : tmp_osd_map.pg_upmap_items) { | |
5461 | if (to_skip.count(pg)) | |
5462 | continue; | |
5463 | if (!only_pools.empty() && !only_pools.count(pg.pool())) | |
5464 | continue; | |
5465 | candidates.push_back(make_pair(pg, um_pair)); | |
5466 | } | |
5467 | if (aggressive) { | |
5468 | // shuffle candidates so they all get equal (in)attention | |
5469 | std::shuffle(candidates.begin(), candidates.end(), get_random_engine(cct, p_seed)); | |
5470 | } | |
5471 | return candidates; | |
5472 | } | |
5473 | ||
31f18b77 FG |
5474 | int OSDMap::get_osds_by_bucket_name(const string &name, set<int> *osds) const |
5475 | { | |
5476 | return crush->get_leaves(name, osds); | |
5477 | } | |
5478 | ||
3efd9988 FG |
5479 | // get pools whose crush rules might reference the given osd |
5480 | void OSDMap::get_pool_ids_by_osd(CephContext *cct, | |
5481 | int osd, | |
5482 | set<int64_t> *pool_ids) const | |
5483 | { | |
11fdf7f2 | 5484 | ceph_assert(pool_ids); |
3efd9988 FG |
5485 | set<int> raw_rules; |
5486 | int r = crush->get_rules_by_osd(osd, &raw_rules); | |
5487 | if (r < 0) { | |
5488 | lderr(cct) << __func__ << " get_rules_by_osd failed: " << cpp_strerror(r) | |
5489 | << dendl; | |
11fdf7f2 | 5490 | ceph_assert(r >= 0); |
3efd9988 FG |
5491 | } |
5492 | set<int> rules; | |
5493 | for (auto &i: raw_rules) { | |
5494 | // exclude any dead rule | |
5495 | if (crush_rule_in_use(i)) { | |
5496 | rules.insert(i); | |
5497 | } | |
5498 | } | |
5499 | for (auto &r: rules) { | |
5500 | get_pool_ids_by_rule(r, pool_ids); | |
5501 | } | |
5502 | } | |
5503 | ||
31f18b77 FG |
5504 | template <typename F> |
5505 | class OSDUtilizationDumper : public CrushTreeDumper::Dumper<F> { | |
5506 | public: | |
5507 | typedef CrushTreeDumper::Dumper<F> Parent; | |
5508 | ||
5509 | OSDUtilizationDumper(const CrushWrapper *crush, const OSDMap *osdmap_, | |
11fdf7f2 | 5510 | const PGMap& pgmap_, bool tree_, |
9f95a23c | 5511 | const string& filter) : |
c07f9fc5 | 5512 | Parent(crush, osdmap_->get_pool_names()), |
31f18b77 | 5513 | osdmap(osdmap_), |
11fdf7f2 | 5514 | pgmap(pgmap_), |
31f18b77 | 5515 | tree(tree_), |
31f18b77 FG |
5516 | min_var(-1), |
5517 | max_var(-1), | |
5518 | stddev(0), | |
5519 | sum(0) { | |
9f95a23c TL |
5520 | if (osdmap->crush->name_exists(filter)) { |
5521 | // filter by crush node | |
5522 | auto item_id = osdmap->crush->get_item_id(filter); | |
11fdf7f2 TL |
5523 | allowed.insert(item_id); |
5524 | osdmap->crush->get_all_children(item_id, &allowed); | |
9f95a23c TL |
5525 | } else if (osdmap->crush->class_exists(filter)) { |
5526 | // filter by device class | |
5527 | class_id = osdmap->crush->get_class_id(filter); | |
5528 | } else if (auto pool_id = osdmap->lookup_pg_pool_name(filter); | |
5529 | pool_id >= 0) { | |
5530 | // filter by pool | |
5531 | auto crush_rule = osdmap->get_pool_crush_rule(pool_id); | |
5532 | set<int> roots; | |
5533 | osdmap->crush->find_takes_by_rule(crush_rule, &roots); | |
5534 | allowed = roots; | |
5535 | for (auto r : roots) | |
5536 | osdmap->crush->get_all_children(r, &allowed); | |
11fdf7f2 TL |
5537 | } |
5538 | average_util = average_utilization(); | |
31f18b77 FG |
5539 | } |
5540 | ||
5541 | protected: | |
11fdf7f2 TL |
5542 | |
5543 | bool should_dump(int id) const { | |
5544 | if (!allowed.empty() && !allowed.count(id)) // filter by name | |
5545 | return false; | |
9f95a23c TL |
5546 | if (id >= 0 && class_id >= 0) { |
5547 | auto item_class_id = osdmap->crush->get_item_class_id(id); | |
5548 | if (item_class_id < 0 || // not bound to a class yet | |
5549 | item_class_id != class_id) // or already bound to a different class | |
11fdf7f2 TL |
5550 | return false; |
5551 | } | |
5552 | return true; | |
5553 | } | |
5554 | ||
5555 | set<int> get_dumped_osds() { | |
9f95a23c | 5556 | if (allowed.empty() && class_id < 0) { |
11fdf7f2 TL |
5557 | // old way, all |
5558 | return {}; | |
5559 | } | |
5560 | return dumped_osds; | |
5561 | } | |
5562 | ||
31f18b77 FG |
5563 | void dump_stray(F *f) { |
5564 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
5565 | if (osdmap->exists(i) && !this->is_touched(i)) | |
c07f9fc5 | 5566 | dump_item(CrushTreeDumper::Item(i, 0, 0, 0), f); |
31f18b77 FG |
5567 | } |
5568 | } | |
5569 | ||
5570 | void dump_item(const CrushTreeDumper::Item &qi, F *f) override { | |
f67539c2 | 5571 | if (!tree && (qi.is_bucket() || dumped_osds.count(qi.id))) |
31f18b77 | 5572 | return; |
11fdf7f2 TL |
5573 | if (!should_dump(qi.id)) |
5574 | return; | |
31f18b77 | 5575 | |
11fdf7f2 TL |
5576 | if (!qi.is_bucket()) |
5577 | dumped_osds.insert(qi.id); | |
31f18b77 | 5578 | float reweight = qi.is_bucket() ? -1 : osdmap->get_weightf(qi.id); |
11fdf7f2 TL |
5579 | int64_t kb = 0, kb_used = 0, kb_used_data = 0, kb_used_omap = 0, |
5580 | kb_used_meta = 0, kb_avail = 0; | |
31f18b77 | 5581 | double util = 0; |
11fdf7f2 TL |
5582 | if (get_bucket_utilization(qi.id, &kb, &kb_used, &kb_used_data, |
5583 | &kb_used_omap, &kb_used_meta, &kb_avail)) | |
31f18b77 FG |
5584 | if (kb_used && kb) |
5585 | util = 100.0 * (double)kb_used / (double)kb; | |
5586 | ||
5587 | double var = 1.0; | |
5588 | if (average_util) | |
5589 | var = util / average_util; | |
5590 | ||
11fdf7f2 | 5591 | size_t num_pgs = qi.is_bucket() ? 0 : pgmap.get_num_pg_by_osd(qi.id); |
31f18b77 | 5592 | |
11fdf7f2 TL |
5593 | dump_item(qi, reweight, kb, kb_used, |
5594 | kb_used_data, kb_used_omap, kb_used_meta, | |
5595 | kb_avail, util, var, num_pgs, f); | |
31f18b77 FG |
5596 | |
5597 | if (!qi.is_bucket() && reweight > 0) { | |
5598 | if (min_var < 0 || var < min_var) | |
5599 | min_var = var; | |
5600 | if (max_var < 0 || var > max_var) | |
5601 | max_var = var; | |
5602 | ||
5603 | double dev = util - average_util; | |
5604 | dev *= dev; | |
5605 | stddev += reweight * dev; | |
5606 | sum += reweight; | |
5607 | } | |
5608 | } | |
5609 | ||
5610 | virtual void dump_item(const CrushTreeDumper::Item &qi, | |
5611 | float &reweight, | |
5612 | int64_t kb, | |
5613 | int64_t kb_used, | |
11fdf7f2 TL |
5614 | int64_t kb_used_data, |
5615 | int64_t kb_used_omap, | |
5616 | int64_t kb_used_meta, | |
31f18b77 FG |
5617 | int64_t kb_avail, |
5618 | double& util, | |
5619 | double& var, | |
5620 | const size_t num_pgs, | |
5621 | F *f) = 0; | |
5622 | ||
5623 | double dev() { | |
5624 | return sum > 0 ? sqrt(stddev / sum) : 0; | |
5625 | } | |
5626 | ||
5627 | double average_utilization() { | |
5628 | int64_t kb = 0, kb_used = 0; | |
5629 | for (int i = 0; i < osdmap->get_max_osd(); i++) { | |
11fdf7f2 TL |
5630 | if (!osdmap->exists(i) || |
5631 | osdmap->get_weight(i) == 0 || | |
5632 | !should_dump(i)) | |
31f18b77 | 5633 | continue; |
11fdf7f2 TL |
5634 | int64_t kb_i, kb_used_i, kb_used_data_i, kb_used_omap_i, kb_used_meta_i, |
5635 | kb_avail_i; | |
5636 | if (get_osd_utilization(i, &kb_i, &kb_used_i, &kb_used_data_i, | |
5637 | &kb_used_omap_i, &kb_used_meta_i, &kb_avail_i)) { | |
31f18b77 FG |
5638 | kb += kb_i; |
5639 | kb_used += kb_used_i; | |
5640 | } | |
5641 | } | |
5642 | return kb > 0 ? 100.0 * (double)kb_used / (double)kb : 0; | |
5643 | } | |
5644 | ||
5645 | bool get_osd_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5646 | int64_t* kb_used_data, |
5647 | int64_t* kb_used_omap, | |
5648 | int64_t* kb_used_meta, | |
31f18b77 | 5649 | int64_t* kb_avail) const { |
11fdf7f2 | 5650 | const osd_stat_t *p = pgmap.get_osd_stat(id); |
31f18b77 | 5651 | if (!p) return false; |
11fdf7f2 TL |
5652 | *kb = p->statfs.kb(); |
5653 | *kb_used = p->statfs.kb_used_raw(); | |
5654 | *kb_used_data = p->statfs.kb_used_data(); | |
5655 | *kb_used_omap = p->statfs.kb_used_omap(); | |
5656 | *kb_used_meta = p->statfs.kb_used_internal_metadata(); | |
5657 | *kb_avail = p->statfs.kb_avail(); | |
5658 | ||
f67539c2 | 5659 | return true; |
31f18b77 FG |
5660 | } |
5661 | ||
5662 | bool get_bucket_utilization(int id, int64_t* kb, int64_t* kb_used, | |
11fdf7f2 TL |
5663 | int64_t* kb_used_data, |
5664 | int64_t* kb_used_omap, | |
5665 | int64_t* kb_used_meta, | |
31f18b77 FG |
5666 | int64_t* kb_avail) const { |
5667 | if (id >= 0) { | |
11fdf7f2 | 5668 | if (osdmap->is_out(id) || !should_dump(id)) { |
31f18b77 FG |
5669 | *kb = 0; |
5670 | *kb_used = 0; | |
11fdf7f2 TL |
5671 | *kb_used_data = 0; |
5672 | *kb_used_omap = 0; | |
5673 | *kb_used_meta = 0; | |
31f18b77 FG |
5674 | *kb_avail = 0; |
5675 | return true; | |
5676 | } | |
11fdf7f2 TL |
5677 | return get_osd_utilization(id, kb, kb_used, kb_used_data, |
5678 | kb_used_omap, kb_used_meta, kb_avail); | |
31f18b77 FG |
5679 | } |
5680 | ||
5681 | *kb = 0; | |
5682 | *kb_used = 0; | |
11fdf7f2 TL |
5683 | *kb_used_data = 0; |
5684 | *kb_used_omap = 0; | |
5685 | *kb_used_meta = 0; | |
31f18b77 FG |
5686 | *kb_avail = 0; |
5687 | ||
5688 | for (int k = osdmap->crush->get_bucket_size(id) - 1; k >= 0; k--) { | |
5689 | int item = osdmap->crush->get_bucket_item(id, k); | |
11fdf7f2 TL |
5690 | int64_t kb_i = 0, kb_used_i = 0, kb_used_data_i = 0, |
5691 | kb_used_omap_i = 0, kb_used_meta_i = 0, kb_avail_i = 0; | |
5692 | if (!get_bucket_utilization(item, &kb_i, &kb_used_i, | |
5693 | &kb_used_data_i, &kb_used_omap_i, | |
5694 | &kb_used_meta_i, &kb_avail_i)) | |
31f18b77 FG |
5695 | return false; |
5696 | *kb += kb_i; | |
5697 | *kb_used += kb_used_i; | |
11fdf7f2 TL |
5698 | *kb_used_data += kb_used_data_i; |
5699 | *kb_used_omap += kb_used_omap_i; | |
5700 | *kb_used_meta += kb_used_meta_i; | |
31f18b77 FG |
5701 | *kb_avail += kb_avail_i; |
5702 | } | |
f67539c2 | 5703 | return true; |
31f18b77 FG |
5704 | } |
5705 | ||
5706 | protected: | |
5707 | const OSDMap *osdmap; | |
11fdf7f2 | 5708 | const PGMap& pgmap; |
31f18b77 FG |
5709 | bool tree; |
5710 | double average_util; | |
5711 | double min_var; | |
5712 | double max_var; | |
5713 | double stddev; | |
5714 | double sum; | |
9f95a23c | 5715 | int class_id = -1; |
11fdf7f2 TL |
5716 | set<int> allowed; |
5717 | set<int> dumped_osds; | |
31f18b77 FG |
5718 | }; |
5719 | ||
5720 | ||
5721 | class OSDUtilizationPlainDumper : public OSDUtilizationDumper<TextTable> { | |
5722 | public: | |
5723 | typedef OSDUtilizationDumper<TextTable> Parent; | |
5724 | ||
5725 | OSDUtilizationPlainDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 5726 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
5727 | const string& filter) : |
5728 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
5729 | |
5730 | void dump(TextTable *tbl) { | |
5731 | tbl->define_column("ID", TextTable::LEFT, TextTable::RIGHT); | |
224ce89b | 5732 | tbl->define_column("CLASS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5733 | tbl->define_column("WEIGHT", TextTable::LEFT, TextTable::RIGHT); |
5734 | tbl->define_column("REWEIGHT", TextTable::LEFT, TextTable::RIGHT); | |
5735 | tbl->define_column("SIZE", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 TL |
5736 | tbl->define_column("RAW USE", TextTable::LEFT, TextTable::RIGHT); |
5737 | tbl->define_column("DATA", TextTable::LEFT, TextTable::RIGHT); | |
5738 | tbl->define_column("OMAP", TextTable::LEFT, TextTable::RIGHT); | |
5739 | tbl->define_column("META", TextTable::LEFT, TextTable::RIGHT); | |
31f18b77 FG |
5740 | tbl->define_column("AVAIL", TextTable::LEFT, TextTable::RIGHT); |
5741 | tbl->define_column("%USE", TextTable::LEFT, TextTable::RIGHT); | |
5742 | tbl->define_column("VAR", TextTable::LEFT, TextTable::RIGHT); | |
5743 | tbl->define_column("PGS", TextTable::LEFT, TextTable::RIGHT); | |
11fdf7f2 | 5744 | tbl->define_column("STATUS", TextTable::LEFT, TextTable::RIGHT); |
31f18b77 FG |
5745 | if (tree) |
5746 | tbl->define_column("TYPE NAME", TextTable::LEFT, TextTable::LEFT); | |
5747 | ||
5748 | Parent::dump(tbl); | |
5749 | ||
5750 | dump_stray(tbl); | |
5751 | ||
11fdf7f2 | 5752 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
224ce89b WB |
5753 | *tbl << "" |
5754 | << "" | |
5755 | << "" << "TOTAL" | |
11fdf7f2 TL |
5756 | << byte_u_t(sum.statfs.total) |
5757 | << byte_u_t(sum.statfs.get_used_raw()) | |
5758 | << byte_u_t(sum.statfs.allocated) | |
5759 | << byte_u_t(sum.statfs.omap_allocated) | |
5760 | << byte_u_t(sum.statfs.internal_metadata) | |
5761 | << byte_u_t(sum.statfs.available) | |
31f18b77 FG |
5762 | << lowprecision_t(average_util) |
5763 | << "" | |
5764 | << TextTable::endrow; | |
5765 | } | |
5766 | ||
5767 | protected: | |
5768 | struct lowprecision_t { | |
5769 | float v; | |
5770 | explicit lowprecision_t(float _v) : v(_v) {} | |
5771 | }; | |
5772 | friend std::ostream &operator<<(ostream& out, const lowprecision_t& v); | |
5773 | ||
5774 | using OSDUtilizationDumper<TextTable>::dump_item; | |
5775 | void dump_item(const CrushTreeDumper::Item &qi, | |
5776 | float &reweight, | |
5777 | int64_t kb, | |
5778 | int64_t kb_used, | |
11fdf7f2 TL |
5779 | int64_t kb_used_data, |
5780 | int64_t kb_used_omap, | |
5781 | int64_t kb_used_meta, | |
31f18b77 FG |
5782 | int64_t kb_avail, |
5783 | double& util, | |
5784 | double& var, | |
5785 | const size_t num_pgs, | |
5786 | TextTable *tbl) override { | |
224ce89b WB |
5787 | const char *c = crush->get_item_class(qi.id); |
5788 | if (!c) | |
5789 | c = ""; | |
31f18b77 | 5790 | *tbl << qi.id |
224ce89b | 5791 | << c |
31f18b77 FG |
5792 | << weightf_t(qi.weight) |
5793 | << weightf_t(reweight) | |
1adf2230 AA |
5794 | << byte_u_t(kb << 10) |
5795 | << byte_u_t(kb_used << 10) | |
11fdf7f2 TL |
5796 | << byte_u_t(kb_used_data << 10) |
5797 | << byte_u_t(kb_used_omap << 10) | |
5798 | << byte_u_t(kb_used_meta << 10) | |
1adf2230 | 5799 | << byte_u_t(kb_avail << 10) |
31f18b77 FG |
5800 | << lowprecision_t(util) |
5801 | << lowprecision_t(var); | |
5802 | ||
5803 | if (qi.is_bucket()) { | |
5804 | *tbl << "-"; | |
11fdf7f2 | 5805 | *tbl << ""; |
31f18b77 FG |
5806 | } else { |
5807 | *tbl << num_pgs; | |
11fdf7f2 TL |
5808 | if (osdmap->is_up(qi.id)) { |
5809 | *tbl << "up"; | |
5810 | } else if (osdmap->is_destroyed(qi.id)) { | |
5811 | *tbl << "destroyed"; | |
5812 | } else { | |
5813 | *tbl << "down"; | |
5814 | } | |
31f18b77 FG |
5815 | } |
5816 | ||
5817 | if (tree) { | |
5818 | ostringstream name; | |
5819 | for (int k = 0; k < qi.depth; k++) | |
5820 | name << " "; | |
5821 | if (qi.is_bucket()) { | |
5822 | int type = crush->get_bucket_type(qi.id); | |
5823 | name << crush->get_type_name(type) << " " | |
5824 | << crush->get_item_name(qi.id); | |
5825 | } else { | |
5826 | name << "osd." << qi.id; | |
5827 | } | |
5828 | *tbl << name.str(); | |
5829 | } | |
5830 | ||
5831 | *tbl << TextTable::endrow; | |
5832 | } | |
5833 | ||
5834 | public: | |
5835 | string summary() { | |
5836 | ostringstream out; | |
5837 | out << "MIN/MAX VAR: " << lowprecision_t(min_var) | |
5838 | << "/" << lowprecision_t(max_var) << " " | |
5839 | << "STDDEV: " << lowprecision_t(dev()); | |
5840 | return out.str(); | |
5841 | } | |
5842 | }; | |
5843 | ||
5844 | ostream& operator<<(ostream& out, | |
5845 | const OSDUtilizationPlainDumper::lowprecision_t& v) | |
5846 | { | |
5847 | if (v.v < -0.01) { | |
5848 | return out << "-"; | |
5849 | } else if (v.v < 0.001) { | |
5850 | return out << "0"; | |
5851 | } else { | |
5852 | std::streamsize p = out.precision(); | |
5853 | return out << std::fixed << std::setprecision(2) << v.v << std::setprecision(p); | |
5854 | } | |
5855 | } | |
5856 | ||
5857 | class OSDUtilizationFormatDumper : public OSDUtilizationDumper<Formatter> { | |
5858 | public: | |
5859 | typedef OSDUtilizationDumper<Formatter> Parent; | |
5860 | ||
5861 | OSDUtilizationFormatDumper(const CrushWrapper *crush, const OSDMap *osdmap, | |
11fdf7f2 | 5862 | const PGMap& pgmap, bool tree, |
9f95a23c TL |
5863 | const string& filter) : |
5864 | Parent(crush, osdmap, pgmap, tree, filter) {} | |
31f18b77 FG |
5865 | |
5866 | void dump(Formatter *f) { | |
5867 | f->open_array_section("nodes"); | |
5868 | Parent::dump(f); | |
5869 | f->close_section(); | |
5870 | ||
5871 | f->open_array_section("stray"); | |
5872 | dump_stray(f); | |
5873 | f->close_section(); | |
5874 | } | |
5875 | ||
5876 | protected: | |
5877 | using OSDUtilizationDumper<Formatter>::dump_item; | |
5878 | void dump_item(const CrushTreeDumper::Item &qi, | |
11fdf7f2 TL |
5879 | float &reweight, |
5880 | int64_t kb, | |
5881 | int64_t kb_used, | |
5882 | int64_t kb_used_data, | |
5883 | int64_t kb_used_omap, | |
5884 | int64_t kb_used_meta, | |
5885 | int64_t kb_avail, | |
5886 | double& util, | |
5887 | double& var, | |
5888 | const size_t num_pgs, | |
5889 | Formatter *f) override { | |
31f18b77 | 5890 | f->open_object_section("item"); |
c07f9fc5 | 5891 | CrushTreeDumper::dump_item_fields(crush, weight_set_names, qi, f); |
31f18b77 FG |
5892 | f->dump_float("reweight", reweight); |
5893 | f->dump_int("kb", kb); | |
5894 | f->dump_int("kb_used", kb_used); | |
11fdf7f2 TL |
5895 | f->dump_int("kb_used_data", kb_used_data); |
5896 | f->dump_int("kb_used_omap", kb_used_omap); | |
5897 | f->dump_int("kb_used_meta", kb_used_meta); | |
31f18b77 FG |
5898 | f->dump_int("kb_avail", kb_avail); |
5899 | f->dump_float("utilization", util); | |
5900 | f->dump_float("var", var); | |
5901 | f->dump_unsigned("pgs", num_pgs); | |
11fdf7f2 TL |
5902 | if (!qi.is_bucket()) { |
5903 | if (osdmap->is_up(qi.id)) { | |
5904 | f->dump_string("status", "up"); | |
5905 | } else if (osdmap->is_destroyed(qi.id)) { | |
5906 | f->dump_string("status", "destroyed"); | |
5907 | } else { | |
5908 | f->dump_string("status", "down"); | |
5909 | } | |
5910 | } | |
31f18b77 FG |
5911 | CrushTreeDumper::dump_bucket_children(crush, qi, f); |
5912 | f->close_section(); | |
5913 | } | |
5914 | ||
5915 | public: | |
5916 | void summary(Formatter *f) { | |
5917 | f->open_object_section("summary"); | |
11fdf7f2 TL |
5918 | auto sum = pgmap.get_osd_sum(get_dumped_osds()); |
5919 | auto& s = sum.statfs; | |
5920 | ||
5921 | f->dump_int("total_kb", s.kb()); | |
5922 | f->dump_int("total_kb_used", s.kb_used_raw()); | |
5923 | f->dump_int("total_kb_used_data", s.kb_used_data()); | |
5924 | f->dump_int("total_kb_used_omap", s.kb_used_omap()); | |
5925 | f->dump_int("total_kb_used_meta", s.kb_used_internal_metadata()); | |
5926 | f->dump_int("total_kb_avail", s.kb_avail()); | |
31f18b77 FG |
5927 | f->dump_float("average_utilization", average_util); |
5928 | f->dump_float("min_var", min_var); | |
5929 | f->dump_float("max_var", max_var); | |
5930 | f->dump_float("dev", dev()); | |
5931 | f->close_section(); | |
5932 | } | |
5933 | }; | |
5934 | ||
5935 | void print_osd_utilization(const OSDMap& osdmap, | |
11fdf7f2 TL |
5936 | const PGMap& pgmap, |
5937 | ostream& out, | |
5938 | Formatter *f, | |
5939 | bool tree, | |
9f95a23c | 5940 | const string& filter) |
31f18b77 FG |
5941 | { |
5942 | const CrushWrapper *crush = osdmap.crush.get(); | |
5943 | if (f) { | |
5944 | f->open_object_section("df"); | |
9f95a23c | 5945 | OSDUtilizationFormatDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
5946 | d.dump(f); |
5947 | d.summary(f); | |
5948 | f->close_section(); | |
5949 | f->flush(out); | |
5950 | } else { | |
9f95a23c | 5951 | OSDUtilizationPlainDumper d(crush, &osdmap, pgmap, tree, filter); |
31f18b77 FG |
5952 | TextTable tbl; |
5953 | d.dump(&tbl); | |
5954 | out << tbl << d.summary() << "\n"; | |
5955 | } | |
5956 | } | |
224ce89b | 5957 | |
92f5a8d4 TL |
5958 | void OSDMap::check_health(CephContext *cct, |
5959 | health_check_map_t *checks) const | |
224ce89b WB |
5960 | { |
5961 | int num_osds = get_num_osds(); | |
5962 | ||
5963 | // OSD_DOWN | |
5964 | // OSD_$subtree_DOWN | |
5965 | // OSD_ORPHAN | |
5966 | if (num_osds >= 0) { | |
5967 | int num_in_osds = 0; | |
5968 | int num_down_in_osds = 0; | |
5969 | set<int> osds; | |
5970 | set<int> down_in_osds; | |
5971 | set<int> up_in_osds; | |
5972 | set<int> subtree_up; | |
5973 | unordered_map<int, set<int> > subtree_type_down; | |
5974 | unordered_map<int, int> num_osds_subtree; | |
5975 | int max_type = crush->get_max_type_id(); | |
5976 | ||
5977 | for (int i = 0; i < get_max_osd(); i++) { | |
5978 | if (!exists(i)) { | |
5979 | if (crush->item_exists(i)) { | |
5980 | osds.insert(i); | |
5981 | } | |
5982 | continue; | |
5983 | } | |
f67539c2 | 5984 | if (is_out(i) || (osd_state[i] & CEPH_OSD_NEW)) |
224ce89b WB |
5985 | continue; |
5986 | ++num_in_osds; | |
5987 | if (down_in_osds.count(i) || up_in_osds.count(i)) | |
5988 | continue; | |
5989 | if (!is_up(i)) { | |
5990 | down_in_osds.insert(i); | |
5991 | int parent_id = 0; | |
5992 | int current = i; | |
5993 | for (int type = 0; type <= max_type; type++) { | |
5994 | if (!crush->get_type_name(type)) | |
5995 | continue; | |
5996 | int r = crush->get_immediate_parent_id(current, &parent_id); | |
5997 | if (r == -ENOENT) | |
5998 | break; | |
5999 | // break early if this parent is already marked as up | |
6000 | if (subtree_up.count(parent_id)) | |
6001 | break; | |
6002 | type = crush->get_bucket_type(parent_id); | |
6003 | if (!subtree_type_is_down( | |
92f5a8d4 | 6004 | cct, parent_id, type, |
224ce89b WB |
6005 | &down_in_osds, &up_in_osds, &subtree_up, &subtree_type_down)) |
6006 | break; | |
6007 | current = parent_id; | |
6008 | } | |
6009 | } | |
6010 | } | |
6011 | ||
6012 | // calculate the number of down osds in each down subtree and | |
6013 | // store it in num_osds_subtree | |
6014 | for (int type = 1; type <= max_type; type++) { | |
6015 | if (!crush->get_type_name(type)) | |
6016 | continue; | |
6017 | for (auto j = subtree_type_down[type].begin(); | |
6018 | j != subtree_type_down[type].end(); | |
6019 | ++j) { | |
6020 | list<int> children; | |
6021 | int num = 0; | |
6022 | int num_children = crush->get_children(*j, &children); | |
6023 | if (num_children == 0) | |
6024 | continue; | |
6025 | for (auto l = children.begin(); l != children.end(); ++l) { | |
6026 | if (*l >= 0) { | |
6027 | ++num; | |
6028 | } else if (num_osds_subtree[*l] > 0) { | |
6029 | num = num + num_osds_subtree[*l]; | |
6030 | } | |
6031 | } | |
6032 | num_osds_subtree[*j] = num; | |
6033 | } | |
6034 | } | |
6035 | num_down_in_osds = down_in_osds.size(); | |
11fdf7f2 | 6036 | ceph_assert(num_down_in_osds <= num_in_osds); |
224ce89b WB |
6037 | if (num_down_in_osds > 0) { |
6038 | // summary of down subtree types and osds | |
6039 | for (int type = max_type; type > 0; type--) { | |
6040 | if (!crush->get_type_name(type)) | |
6041 | continue; | |
6042 | if (subtree_type_down[type].size() > 0) { | |
6043 | ostringstream ss; | |
6044 | ss << subtree_type_down[type].size() << " " | |
6045 | << crush->get_type_name(type); | |
6046 | if (subtree_type_down[type].size() > 1) { | |
6047 | ss << "s"; | |
6048 | } | |
6049 | int sum_down_osds = 0; | |
6050 | for (auto j = subtree_type_down[type].begin(); | |
6051 | j != subtree_type_down[type].end(); | |
6052 | ++j) { | |
6053 | sum_down_osds = sum_down_osds + num_osds_subtree[*j]; | |
6054 | } | |
6055 | ss << " (" << sum_down_osds << " osds) down"; | |
6056 | string err = string("OSD_") + | |
6057 | string(crush->get_type_name(type)) + "_DOWN"; | |
6058 | boost::to_upper(err); | |
9f95a23c TL |
6059 | auto& d = checks->add(err, HEALTH_WARN, ss.str(), |
6060 | subtree_type_down[type].size()); | |
224ce89b WB |
6061 | for (auto j = subtree_type_down[type].rbegin(); |
6062 | j != subtree_type_down[type].rend(); | |
6063 | ++j) { | |
6064 | ostringstream ss; | |
6065 | ss << crush->get_type_name(type); | |
6066 | ss << " "; | |
6067 | ss << crush->get_item_name(*j); | |
6068 | // at the top level, do not print location | |
6069 | if (type != max_type) { | |
6070 | ss << " ("; | |
6071 | ss << crush->get_full_location_ordered_string(*j); | |
6072 | ss << ")"; | |
6073 | } | |
6074 | int num = num_osds_subtree[*j]; | |
6075 | ss << " (" << num << " osds)"; | |
6076 | ss << " is down"; | |
6077 | d.detail.push_back(ss.str()); | |
6078 | } | |
6079 | } | |
6080 | } | |
6081 | ostringstream ss; | |
6082 | ss << down_in_osds.size() << " osds down"; | |
9f95a23c TL |
6083 | auto& d = checks->add("OSD_DOWN", HEALTH_WARN, ss.str(), |
6084 | down_in_osds.size()); | |
224ce89b WB |
6085 | for (auto it = down_in_osds.begin(); it != down_in_osds.end(); ++it) { |
6086 | ostringstream ss; | |
6087 | ss << "osd." << *it << " ("; | |
6088 | ss << crush->get_full_location_ordered_string(*it); | |
6089 | ss << ") is down"; | |
6090 | d.detail.push_back(ss.str()); | |
6091 | } | |
6092 | } | |
6093 | ||
6094 | if (!osds.empty()) { | |
6095 | ostringstream ss; | |
6096 | ss << osds.size() << " osds exist in the crush map but not in the osdmap"; | |
9f95a23c TL |
6097 | auto& d = checks->add("OSD_ORPHAN", HEALTH_WARN, ss.str(), |
6098 | osds.size()); | |
224ce89b WB |
6099 | for (auto osd : osds) { |
6100 | ostringstream ss; | |
6101 | ss << "osd." << osd << " exists in crush map but not in osdmap"; | |
6102 | d.detail.push_back(ss.str()); | |
6103 | } | |
6104 | } | |
6105 | } | |
6106 | ||
eafe8130 TL |
6107 | std::list<std::string> scrub_messages; |
6108 | bool noscrub = false, nodeepscrub = false; | |
6109 | for (const auto &p : pools) { | |
6110 | if (p.second.flags & pg_pool_t::FLAG_NOSCRUB) { | |
6111 | ostringstream ss; | |
6112 | ss << "Pool " << get_pool_name(p.first) << " has noscrub flag"; | |
6113 | scrub_messages.push_back(ss.str()); | |
6114 | noscrub = true; | |
6115 | } | |
6116 | if (p.second.flags & pg_pool_t::FLAG_NODEEP_SCRUB) { | |
6117 | ostringstream ss; | |
6118 | ss << "Pool " << get_pool_name(p.first) << " has nodeep-scrub flag"; | |
6119 | scrub_messages.push_back(ss.str()); | |
6120 | nodeepscrub = true; | |
6121 | } | |
6122 | } | |
6123 | if (noscrub || nodeepscrub) { | |
6124 | string out = ""; | |
6125 | out += noscrub ? string("noscrub") + (nodeepscrub ? ", " : "") : ""; | |
6126 | out += nodeepscrub ? "nodeep-scrub" : ""; | |
6127 | auto& d = checks->add("POOL_SCRUB_FLAGS", HEALTH_OK, | |
9f95a23c | 6128 | "Some pool(s) have the " + out + " flag(s) set", 0); |
eafe8130 TL |
6129 | d.detail.splice(d.detail.end(), scrub_messages); |
6130 | } | |
6131 | ||
224ce89b WB |
6132 | // OSD_OUT_OF_ORDER_FULL |
6133 | { | |
6134 | // An osd could configure failsafe ratio, to something different | |
6135 | // but for now assume it is the same here. | |
92f5a8d4 | 6136 | float fsr = cct->_conf->osd_failsafe_full_ratio; |
224ce89b WB |
6137 | if (fsr > 1.0) fsr /= 100; |
6138 | float fr = get_full_ratio(); | |
6139 | float br = get_backfillfull_ratio(); | |
6140 | float nr = get_nearfull_ratio(); | |
6141 | ||
6142 | list<string> detail; | |
6143 | // These checks correspond to how OSDService::check_full_status() in an OSD | |
6144 | // handles the improper setting of these values. | |
6145 | if (br < nr) { | |
6146 | ostringstream ss; | |
6147 | ss << "backfillfull_ratio (" << br | |
6148 | << ") < nearfull_ratio (" << nr << "), increased"; | |
6149 | detail.push_back(ss.str()); | |
6150 | br = nr; | |
6151 | } | |
6152 | if (fr < br) { | |
6153 | ostringstream ss; | |
6154 | ss << "full_ratio (" << fr << ") < backfillfull_ratio (" << br | |
6155 | << "), increased"; | |
6156 | detail.push_back(ss.str()); | |
6157 | fr = br; | |
6158 | } | |
6159 | if (fsr < fr) { | |
6160 | ostringstream ss; | |
6161 | ss << "osd_failsafe_full_ratio (" << fsr << ") < full_ratio (" << fr | |
6162 | << "), increased"; | |
6163 | detail.push_back(ss.str()); | |
6164 | } | |
6165 | if (!detail.empty()) { | |
6166 | auto& d = checks->add("OSD_OUT_OF_ORDER_FULL", HEALTH_ERR, | |
9f95a23c | 6167 | "full ratio(s) out of order", 0); |
224ce89b WB |
6168 | d.detail.swap(detail); |
6169 | } | |
6170 | } | |
6171 | ||
6172 | // OSD_FULL | |
6173 | // OSD_NEARFULL | |
6174 | // OSD_BACKFILLFULL | |
6175 | // OSD_FAILSAFE_FULL | |
6176 | { | |
6177 | set<int> full, backfillfull, nearfull; | |
6178 | get_full_osd_counts(&full, &backfillfull, &nearfull); | |
6179 | if (full.size()) { | |
6180 | ostringstream ss; | |
6181 | ss << full.size() << " full osd(s)"; | |
9f95a23c | 6182 | auto& d = checks->add("OSD_FULL", HEALTH_ERR, ss.str(), full.size()); |
224ce89b WB |
6183 | for (auto& i: full) { |
6184 | ostringstream ss; | |
6185 | ss << "osd." << i << " is full"; | |
6186 | d.detail.push_back(ss.str()); | |
6187 | } | |
6188 | } | |
6189 | if (backfillfull.size()) { | |
6190 | ostringstream ss; | |
6191 | ss << backfillfull.size() << " backfillfull osd(s)"; | |
9f95a23c TL |
6192 | auto& d = checks->add("OSD_BACKFILLFULL", HEALTH_WARN, ss.str(), |
6193 | backfillfull.size()); | |
224ce89b WB |
6194 | for (auto& i: backfillfull) { |
6195 | ostringstream ss; | |
6196 | ss << "osd." << i << " is backfill full"; | |
6197 | d.detail.push_back(ss.str()); | |
6198 | } | |
6199 | } | |
6200 | if (nearfull.size()) { | |
6201 | ostringstream ss; | |
6202 | ss << nearfull.size() << " nearfull osd(s)"; | |
9f95a23c | 6203 | auto& d = checks->add("OSD_NEARFULL", HEALTH_WARN, ss.str(), nearfull.size()); |
224ce89b WB |
6204 | for (auto& i: nearfull) { |
6205 | ostringstream ss; | |
6206 | ss << "osd." << i << " is near full"; | |
6207 | d.detail.push_back(ss.str()); | |
6208 | } | |
6209 | } | |
6210 | } | |
6211 | ||
6212 | // OSDMAP_FLAGS | |
6213 | { | |
6214 | // warn about flags | |
6215 | uint64_t warn_flags = | |
224ce89b WB |
6216 | CEPH_OSDMAP_PAUSERD | |
6217 | CEPH_OSDMAP_PAUSEWR | | |
6218 | CEPH_OSDMAP_PAUSEREC | | |
6219 | CEPH_OSDMAP_NOUP | | |
6220 | CEPH_OSDMAP_NODOWN | | |
6221 | CEPH_OSDMAP_NOIN | | |
6222 | CEPH_OSDMAP_NOOUT | | |
6223 | CEPH_OSDMAP_NOBACKFILL | | |
6224 | CEPH_OSDMAP_NORECOVER | | |
6225 | CEPH_OSDMAP_NOSCRUB | | |
6226 | CEPH_OSDMAP_NODEEP_SCRUB | | |
6227 | CEPH_OSDMAP_NOTIERAGENT | | |
11fdf7f2 | 6228 | CEPH_OSDMAP_NOSNAPTRIM | |
224ce89b WB |
6229 | CEPH_OSDMAP_NOREBALANCE; |
6230 | if (test_flag(warn_flags)) { | |
6231 | ostringstream ss; | |
9f95a23c TL |
6232 | string s = get_flag_string(get_flags() & warn_flags); |
6233 | ss << s << " flag(s) set"; | |
6234 | checks->add("OSDMAP_FLAGS", HEALTH_WARN, ss.str(), | |
6235 | s.size() /* kludgey but sufficient */); | |
224ce89b WB |
6236 | } |
6237 | } | |
6238 | ||
6239 | // OSD_FLAGS | |
6240 | { | |
6241 | list<string> detail; | |
6242 | const unsigned flags = | |
6243 | CEPH_OSD_NOUP | | |
6244 | CEPH_OSD_NOIN | | |
6245 | CEPH_OSD_NODOWN | | |
6246 | CEPH_OSD_NOOUT; | |
6247 | for (int i = 0; i < max_osd; ++i) { | |
6248 | if (osd_state[i] & flags) { | |
6249 | ostringstream ss; | |
6250 | set<string> states; | |
6251 | OSDMap::calc_state_set(osd_state[i] & flags, states); | |
6252 | ss << "osd." << i << " has flags " << states; | |
6253 | detail.push_back(ss.str()); | |
6254 | } | |
6255 | } | |
81eedcae TL |
6256 | for (auto& i : crush_node_flags) { |
6257 | if (i.second && crush->item_exists(i.first)) { | |
6258 | ostringstream ss; | |
6259 | set<string> states; | |
6260 | OSDMap::calc_state_set(i.second, states); | |
6261 | int t = i.first >= 0 ? 0 : crush->get_bucket_type(i.first); | |
6262 | const char *tn = crush->get_type_name(t); | |
6263 | ss << (tn ? tn : "node") << " " | |
6264 | << crush->get_item_name(i.first) << " has flags " << states; | |
6265 | detail.push_back(ss.str()); | |
6266 | } | |
6267 | } | |
6268 | for (auto& i : device_class_flags) { | |
6269 | const char* class_name = crush->get_class_name(i.first); | |
6270 | if (i.second && class_name) { | |
6271 | ostringstream ss; | |
6272 | set<string> states; | |
6273 | OSDMap::calc_state_set(i.second, states); | |
6274 | ss << "device class '" << class_name << "' has flags " << states; | |
6275 | detail.push_back(ss.str()); | |
6276 | } | |
6277 | } | |
224ce89b WB |
6278 | if (!detail.empty()) { |
6279 | ostringstream ss; | |
81eedcae | 6280 | ss << detail.size() << " OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set"; |
9f95a23c | 6281 | auto& d = checks->add("OSD_FLAGS", HEALTH_WARN, ss.str(), detail.size()); |
224ce89b WB |
6282 | d.detail.swap(detail); |
6283 | } | |
6284 | } | |
6285 | ||
6286 | // OLD_CRUSH_TUNABLES | |
92f5a8d4 | 6287 | if (cct->_conf->mon_warn_on_legacy_crush_tunables) { |
224ce89b | 6288 | string min = crush->get_min_required_version(); |
92f5a8d4 | 6289 | if (min < cct->_conf->mon_crush_min_required_version) { |
224ce89b WB |
6290 | ostringstream ss; |
6291 | ss << "crush map has legacy tunables (require " << min | |
92f5a8d4 | 6292 | << ", min is " << cct->_conf->mon_crush_min_required_version << ")"; |
9f95a23c | 6293 | auto& d = checks->add("OLD_CRUSH_TUNABLES", HEALTH_WARN, ss.str(), 0); |
f67539c2 | 6294 | d.detail.push_back("see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
6295 | } |
6296 | } | |
6297 | ||
6298 | // OLD_CRUSH_STRAW_CALC_VERSION | |
92f5a8d4 | 6299 | if (cct->_conf->mon_warn_on_crush_straw_calc_version_zero) { |
224ce89b WB |
6300 | if (crush->get_straw_calc_version() == 0) { |
6301 | ostringstream ss; | |
6302 | ss << "crush map has straw_calc_version=0"; | |
9f95a23c | 6303 | auto& d = checks->add("OLD_CRUSH_STRAW_CALC_VERSION", HEALTH_WARN, ss.str(), 0); |
224ce89b | 6304 | d.detail.push_back( |
f67539c2 | 6305 | "see http://docs.ceph.com/en/latest/rados/operations/crush-map/#tunables"); |
224ce89b WB |
6306 | } |
6307 | } | |
6308 | ||
6309 | // CACHE_POOL_NO_HIT_SET | |
92f5a8d4 | 6310 | if (cct->_conf->mon_warn_on_cache_pools_without_hit_sets) { |
224ce89b | 6311 | list<string> detail; |
9f95a23c | 6312 | for (auto p = pools.cbegin(); p != pools.cend(); ++p) { |
224ce89b WB |
6313 | const pg_pool_t& info = p->second; |
6314 | if (info.cache_mode_requires_hit_set() && | |
6315 | info.hit_set_params.get_type() == HitSet::TYPE_NONE) { | |
6316 | ostringstream ss; | |
6317 | ss << "pool '" << get_pool_name(p->first) | |
6318 | << "' with cache_mode " << info.get_cache_mode_name() | |
6319 | << " needs hit_set_type to be set but it is not"; | |
6320 | detail.push_back(ss.str()); | |
6321 | } | |
6322 | } | |
6323 | if (!detail.empty()) { | |
6324 | ostringstream ss; | |
6325 | ss << detail.size() << " cache pools are missing hit_sets"; | |
9f95a23c TL |
6326 | auto& d = checks->add("CACHE_POOL_NO_HIT_SET", HEALTH_WARN, ss.str(), |
6327 | detail.size()); | |
224ce89b WB |
6328 | d.detail.swap(detail); |
6329 | } | |
6330 | } | |
6331 | ||
6332 | // OSD_NO_SORTBITWISE | |
11fdf7f2 | 6333 | if (!test_flag(CEPH_OSDMAP_SORTBITWISE)) { |
224ce89b | 6334 | ostringstream ss; |
11fdf7f2 | 6335 | ss << "'sortbitwise' flag is not set"; |
9f95a23c | 6336 | checks->add("OSD_NO_SORTBITWISE", HEALTH_WARN, ss.str(), 0); |
224ce89b WB |
6337 | } |
6338 | ||
6339 | // OSD_UPGRADE_FINISHED | |
20effc67 TL |
6340 | if (auto require_release = pending_require_osd_release()) { |
6341 | ostringstream ss; | |
6342 | ss << "all OSDs are running " << *require_release << " or later but" | |
6343 | << " require_osd_release < " << *require_release; | |
6344 | auto& d = checks->add("OSD_UPGRADE_FINISHED", HEALTH_WARN, ss.str(), 0); | |
6345 | d.detail.push_back(ss.str()); | |
6346 | } | |
224ce89b | 6347 | |
3efd9988 | 6348 | // POOL_NEARFULL/BACKFILLFULL/FULL |
224ce89b | 6349 | { |
3efd9988 | 6350 | list<string> full_detail, backfillfull_detail, nearfull_detail; |
224ce89b WB |
6351 | for (auto it : get_pools()) { |
6352 | const pg_pool_t &pool = it.second; | |
3efd9988 | 6353 | const string& pool_name = get_pool_name(it.first); |
224ce89b | 6354 | if (pool.has_flag(pg_pool_t::FLAG_FULL)) { |
224ce89b | 6355 | stringstream ss; |
11fdf7f2 | 6356 | if (pool.has_flag(pg_pool_t::FLAG_FULL_QUOTA)) { |
3efd9988 FG |
6357 | // may run out of space too, |
6358 | // but we want EQUOTA taking precedence | |
11fdf7f2 | 6359 | ss << "pool '" << pool_name << "' is full (running out of quota)"; |
3efd9988 FG |
6360 | } else { |
6361 | ss << "pool '" << pool_name << "' is full (no space)"; | |
6362 | } | |
6363 | full_detail.push_back(ss.str()); | |
6364 | } else if (pool.has_flag(pg_pool_t::FLAG_BACKFILLFULL)) { | |
6365 | stringstream ss; | |
6366 | ss << "pool '" << pool_name << "' is backfillfull"; | |
6367 | backfillfull_detail.push_back(ss.str()); | |
6368 | } else if (pool.has_flag(pg_pool_t::FLAG_NEARFULL)) { | |
6369 | stringstream ss; | |
6370 | ss << "pool '" << pool_name << "' is nearfull"; | |
6371 | nearfull_detail.push_back(ss.str()); | |
224ce89b WB |
6372 | } |
6373 | } | |
3efd9988 | 6374 | if (!full_detail.empty()) { |
224ce89b | 6375 | ostringstream ss; |
3efd9988 | 6376 | ss << full_detail.size() << " pool(s) full"; |
9f95a23c | 6377 | auto& d = checks->add("POOL_FULL", HEALTH_WARN, ss.str(), full_detail.size()); |
3efd9988 FG |
6378 | d.detail.swap(full_detail); |
6379 | } | |
6380 | if (!backfillfull_detail.empty()) { | |
6381 | ostringstream ss; | |
6382 | ss << backfillfull_detail.size() << " pool(s) backfillfull"; | |
9f95a23c TL |
6383 | auto& d = checks->add("POOL_BACKFILLFULL", HEALTH_WARN, ss.str(), |
6384 | backfillfull_detail.size()); | |
3efd9988 FG |
6385 | d.detail.swap(backfillfull_detail); |
6386 | } | |
6387 | if (!nearfull_detail.empty()) { | |
6388 | ostringstream ss; | |
6389 | ss << nearfull_detail.size() << " pool(s) nearfull"; | |
9f95a23c TL |
6390 | auto& d = checks->add("POOL_NEARFULL", HEALTH_WARN, ss.str(), |
6391 | nearfull_detail.size()); | |
3efd9988 | 6392 | d.detail.swap(nearfull_detail); |
224ce89b WB |
6393 | } |
6394 | } | |
92f5a8d4 TL |
6395 | |
6396 | // POOL_PG_NUM_NOT_POWER_OF_TWO | |
6397 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_pg_num_not_power_of_two")) { | |
6398 | list<string> detail; | |
6399 | for (auto it : get_pools()) { | |
6400 | if (!isp2(it.second.get_pg_num_target())) { | |
6401 | ostringstream ss; | |
6402 | ss << "pool '" << get_pool_name(it.first) | |
6403 | << "' pg_num " << it.second.get_pg_num_target() | |
6404 | << " is not a power of two"; | |
6405 | detail.push_back(ss.str()); | |
6406 | } | |
6407 | } | |
6408 | if (!detail.empty()) { | |
6409 | ostringstream ss; | |
6410 | ss << detail.size() << " pool(s) have non-power-of-two pg_num"; | |
6411 | auto& d = checks->add("POOL_PG_NUM_NOT_POWER_OF_TWO", HEALTH_WARN, | |
9f95a23c TL |
6412 | ss.str(), detail.size()); |
6413 | d.detail.swap(detail); | |
6414 | } | |
6415 | } | |
6416 | ||
6417 | // POOL_NO_REDUNDANCY | |
6418 | if (cct->_conf.get_val<bool>("mon_warn_on_pool_no_redundancy")) | |
6419 | { | |
6420 | list<string> detail; | |
6421 | for (auto it : get_pools()) { | |
6422 | if (it.second.get_size() == 1) { | |
6423 | ostringstream ss; | |
6424 | ss << "pool '" << get_pool_name(it.first) | |
6425 | << "' has no replicas configured"; | |
6426 | detail.push_back(ss.str()); | |
6427 | } | |
6428 | } | |
6429 | if (!detail.empty()) { | |
6430 | ostringstream ss; | |
6431 | ss << detail.size() << " pool(s) have no replicas configured"; | |
6432 | auto& d = checks->add("POOL_NO_REDUNDANCY", HEALTH_WARN, | |
6433 | ss.str(), detail.size()); | |
92f5a8d4 TL |
6434 | d.detail.swap(detail); |
6435 | } | |
6436 | } | |
f67539c2 TL |
6437 | |
6438 | // DEGRADED STRETCH MODE | |
6439 | if (cct->_conf.get_val<bool>("mon_warn_on_degraded_stretch_mode")) { | |
6440 | if (recovering_stretch_mode) { | |
6441 | stringstream ss; | |
6442 | ss << "We are recovering stretch mode buckets, only requiring " | |
6443 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
6444 | checks->add("RECOVERING_STRETCH_MODE", HEALTH_WARN, | |
6445 | ss.str(), 0); | |
6446 | } else if (degraded_stretch_mode) { | |
6447 | stringstream ss; | |
6448 | ss << "We are missing stretch mode buckets, only requiring " | |
6449 | << degraded_stretch_mode << " of " << stretch_bucket_count << " buckets to peer" ; | |
6450 | checks->add("DEGRADED_STRETCH_MODE", HEALTH_WARN, | |
6451 | ss.str(), 0); | |
6452 | } | |
6453 | } | |
224ce89b | 6454 | } |
35e4c445 FG |
6455 | |
6456 | int OSDMap::parse_osd_id_list(const vector<string>& ls, set<int> *out, | |
6457 | ostream *ss) const | |
6458 | { | |
6459 | out->clear(); | |
6460 | for (auto i = ls.begin(); i != ls.end(); ++i) { | |
6461 | if (i == ls.begin() && | |
6462 | (*i == "any" || *i == "all" || *i == "*")) { | |
6463 | get_all_osds(*out); | |
6464 | break; | |
6465 | } | |
9f95a23c | 6466 | long osd = TOPNSPC::common::parse_osd_id(i->c_str(), ss); |
35e4c445 FG |
6467 | if (osd < 0) { |
6468 | *ss << "invalid osd id '" << *i << "'"; | |
6469 | return -EINVAL; | |
6470 | } | |
6471 | out->insert(osd); | |
6472 | } | |
6473 | return 0; | |
6474 | } | |
11fdf7f2 TL |
6475 | |
6476 | void OSDMap::get_random_up_osds_by_subtree(int n, // whoami | |
6477 | string &subtree, | |
6478 | int limit, // how many | |
6479 | set<int> skip, | |
6480 | set<int> *want) const { | |
6481 | if (limit <= 0) | |
6482 | return; | |
6483 | int subtree_type = crush->get_type_id(subtree); | |
6484 | if (subtree_type < 1) | |
6485 | return; | |
6486 | vector<int> subtrees; | |
6487 | crush->get_subtree_of_type(subtree_type, &subtrees); | |
6488 | std::random_device rd; | |
6489 | std::default_random_engine rng{rd()}; | |
6490 | std::shuffle(subtrees.begin(), subtrees.end(), rng); | |
6491 | for (auto s : subtrees) { | |
6492 | if (limit <= 0) | |
6493 | break; | |
6494 | if (crush->subtree_contains(s, n)) | |
6495 | continue; | |
6496 | vector<int> osds; | |
6497 | crush->get_children_of_type(s, 0, &osds); | |
6498 | if (osds.empty()) | |
6499 | continue; | |
6500 | vector<int> up_osds; | |
6501 | for (auto o : osds) { | |
6502 | if (is_up(o) && !skip.count(o)) | |
6503 | up_osds.push_back(o); | |
6504 | } | |
6505 | if (up_osds.empty()) | |
6506 | continue; | |
6507 | auto it = up_osds.begin(); | |
6508 | std::advance(it, (n % up_osds.size())); | |
6509 | want->insert(*it); | |
6510 | --limit; | |
6511 | } | |
6512 | } | |
6513 | ||
6514 | float OSDMap::pool_raw_used_rate(int64_t poolid) const | |
6515 | { | |
6516 | const pg_pool_t *pool = get_pg_pool(poolid); | |
6517 | assert(pool != nullptr); | |
6518 | ||
6519 | switch (pool->get_type()) { | |
6520 | case pg_pool_t::TYPE_REPLICATED: | |
6521 | return pool->get_size(); | |
11fdf7f2 TL |
6522 | case pg_pool_t::TYPE_ERASURE: |
6523 | { | |
6524 | auto& ecp = | |
6525 | get_erasure_code_profile(pool->erasure_code_profile); | |
6526 | auto pm = ecp.find("m"); | |
6527 | auto pk = ecp.find("k"); | |
6528 | if (pm != ecp.end() && pk != ecp.end()) { | |
6529 | int k = atoi(pk->second.c_str()); | |
6530 | int m = atoi(pm->second.c_str()); | |
6531 | int mk = m + k; | |
6532 | ceph_assert(mk != 0); | |
6533 | ceph_assert(k != 0); | |
6534 | return (float)mk / k; | |
6535 | } else { | |
6536 | return 0.0; | |
6537 | } | |
6538 | } | |
6539 | break; | |
6540 | default: | |
6541 | ceph_abort_msg("unrecognized pool type"); | |
6542 | } | |
6543 | } | |
81eedcae TL |
6544 | |
6545 | unsigned OSDMap::get_osd_crush_node_flags(int osd) const | |
6546 | { | |
6547 | unsigned flags = 0; | |
6548 | if (!crush_node_flags.empty()) { | |
6549 | // the map will contain type -> name | |
6550 | std::map<std::string,std::string> ploc = crush->get_full_location(osd); | |
6551 | for (auto& i : ploc) { | |
6552 | int id = crush->get_item_id(i.second); | |
6553 | auto p = crush_node_flags.find(id); | |
6554 | if (p != crush_node_flags.end()) { | |
6555 | flags |= p->second; | |
6556 | } | |
6557 | } | |
6558 | } | |
6559 | return flags; | |
6560 | } | |
6561 | ||
6562 | unsigned OSDMap::get_crush_node_flags(int id) const | |
6563 | { | |
6564 | unsigned flags = 0; | |
6565 | auto it = crush_node_flags.find(id); | |
6566 | if (it != crush_node_flags.end()) | |
6567 | flags = it->second; | |
6568 | return flags; | |
6569 | } | |
6570 | ||
6571 | unsigned OSDMap::get_device_class_flags(int id) const | |
6572 | { | |
6573 | unsigned flags = 0; | |
6574 | auto it = device_class_flags.find(id); | |
6575 | if (it != device_class_flags.end()) | |
6576 | flags = it->second; | |
6577 | return flags; | |
6578 | } | |
20effc67 TL |
6579 | |
6580 | std::optional<std::string> OSDMap::pending_require_osd_release() const | |
6581 | { | |
6582 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_QUINCY) && | |
6583 | require_osd_release < ceph_release_t::quincy) { | |
6584 | return "quincy"; | |
6585 | } | |
6586 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_PACIFIC) && | |
6587 | require_osd_release < ceph_release_t::pacific) { | |
6588 | return "pacific"; | |
6589 | } | |
6590 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_OCTOPUS) && | |
6591 | require_osd_release < ceph_release_t::octopus) { | |
6592 | return "octopus"; | |
6593 | } | |
6594 | if (HAVE_FEATURE(get_up_osd_features(), SERVER_NAUTILUS) && | |
6595 | require_osd_release < ceph_release_t::nautilus) { | |
6596 | return "nautilus"; | |
6597 | } | |
6598 | ||
6599 | return std::nullopt; | |
6600 | } |