]>
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 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #include "include/compat.h" | |
16 | #include "mdstypes.h" | |
17 | ||
91327a77 | 18 | #include "mon/MonClient.h" |
7c673cae FG |
19 | #include "MDBalancer.h" |
20 | #include "MDSRank.h" | |
7c673cae FG |
21 | #include "MDSMap.h" |
22 | #include "CInode.h" | |
23 | #include "CDir.h" | |
24 | #include "MDCache.h" | |
25 | #include "Migrator.h" | |
26 | #include "Mantle.h" | |
27 | ||
28 | #include "include/Context.h" | |
29 | #include "msg/Messenger.h" | |
7c673cae FG |
30 | |
31 | #include <fstream> | |
7c673cae FG |
32 | #include <vector> |
33 | #include <map> | |
20effc67 TL |
34 | |
35 | using namespace std; | |
7c673cae FG |
36 | |
37 | #include "common/config.h" | |
38 | #include "common/errno.h" | |
39 | ||
1e59de90 TL |
40 | /* Note, by default debug_mds_balancer is 1/5. For debug messages 1<lvl<=5, |
41 | * should_gather (below) will be true; so, debug_mds will be ignored even if | |
42 | * set to 20/20. For this reason, some messages may not appear in the log. | |
43 | * Increase both debug levels to get expected output! | |
44 | */ | |
7c673cae | 45 | #define dout_context g_ceph_context |
7c673cae | 46 | #undef dout_prefix |
9f95a23c | 47 | #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".bal " << __func__ << " " |
7c673cae FG |
48 | #undef dout |
49 | #define dout(lvl) \ | |
50 | do {\ | |
51 | auto subsys = ceph_subsys_mds;\ | |
52 | if ((dout_context)->_conf->subsys.should_gather(ceph_subsys_mds_balancer, lvl)) {\ | |
53 | subsys = ceph_subsys_mds_balancer;\ | |
54 | }\ | |
11fdf7f2 | 55 | dout_impl(dout_context, ceph::dout::need_dynamic(subsys), lvl) dout_prefix |
7c673cae FG |
56 | #undef dendl |
57 | #define dendl dendl_impl; } while (0) | |
58 | ||
59 | ||
60 | #define MIN_LOAD 50 // ?? | |
61 | #define MIN_REEXPORT 5 // will automatically reexport | |
62 | #define MIN_OFFLOAD 10 // point at which i stop trying, close enough | |
63 | ||
64 | ||
9f95a23c | 65 | int MDBalancer::proc_message(const cref_t<Message> &m) |
7c673cae FG |
66 | { |
67 | switch (m->get_type()) { | |
68 | ||
69 | case MSG_MDS_HEARTBEAT: | |
9f95a23c | 70 | handle_heartbeat(ref_cast<MHeartbeat>(m)); |
7c673cae FG |
71 | break; |
72 | ||
73 | default: | |
b32b8144 | 74 | derr << " balancer unknown message " << m->get_type() << dendl_impl; |
11fdf7f2 | 75 | ceph_abort_msg("balancer unknown message"); |
7c673cae FG |
76 | } |
77 | ||
78 | return 0; | |
79 | } | |
80 | ||
91327a77 AA |
81 | MDBalancer::MDBalancer(MDSRank *m, Messenger *msgr, MonClient *monc) : |
82 | mds(m), messenger(msgr), mon_client(monc) | |
83 | { | |
11fdf7f2 TL |
84 | bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); |
85 | bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); | |
91327a77 AA |
86 | } |
87 | ||
92f5a8d4 | 88 | void MDBalancer::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mds_map) |
91327a77 | 89 | { |
1e59de90 | 90 | if (changed.count("mds_bal_fragment_dirs")) { |
11fdf7f2 | 91 | bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); |
1e59de90 TL |
92 | } |
93 | if (changed.count("mds_bal_fragment_interval")) { | |
11fdf7f2 | 94 | bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); |
1e59de90 TL |
95 | } |
96 | } | |
97 | ||
98 | bool MDBalancer::test_rank_mask(mds_rank_t rank) | |
99 | { | |
100 | return mds->mdsmap->get_bal_rank_mask_bitset().test(rank); | |
91327a77 AA |
101 | } |
102 | ||
7c673cae FG |
103 | void MDBalancer::handle_export_pins(void) |
104 | { | |
f67539c2 TL |
105 | const mds_rank_t max_mds = mds->mdsmap->get_max_mds(); |
106 | auto mdcache = mds->mdcache; | |
107 | ||
108 | auto &q = mdcache->export_pin_queue; | |
7c673cae FG |
109 | auto it = q.begin(); |
110 | dout(20) << "export_pin_queue size=" << q.size() << dendl; | |
111 | while (it != q.end()) { | |
31f18b77 FG |
112 | auto cur = it++; |
113 | CInode *in = *cur; | |
11fdf7f2 | 114 | ceph_assert(in->is_dir()); |
f6b5b4d7 | 115 | |
31f18b77 | 116 | mds_rank_t export_pin = in->get_export_pin(false); |
f67539c2 TL |
117 | in->check_pin_policy(export_pin); |
118 | ||
119 | if (export_pin >= max_mds) { | |
f6b5b4d7 | 120 | dout(20) << " delay export_pin=" << export_pin << " on " << *in << dendl; |
eafe8130 TL |
121 | in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); |
122 | q.erase(cur); | |
123 | ||
124 | in->state_set(CInode::STATE_DELAYEDEXPORTPIN); | |
f67539c2 | 125 | mdcache->export_pin_delayed_queue.insert(in); |
eafe8130 TL |
126 | continue; |
127 | } | |
31f18b77 | 128 | |
f67539c2 TL |
129 | dout(20) << " executing export_pin=" << export_pin << " on " << *in << dendl; |
130 | unsigned min_frag_bits = 0; | |
131 | mds_rank_t target = MDS_RANK_NONE; | |
132 | if (export_pin >= 0) | |
133 | target = export_pin; | |
134 | else if (export_pin == MDS_RANK_EPHEMERAL_RAND) | |
135 | target = mdcache->hash_into_rank_bucket(in->ino()); | |
136 | else if (export_pin == MDS_RANK_EPHEMERAL_DIST) | |
137 | min_frag_bits = mdcache->get_ephemeral_dist_frag_bits(); | |
138 | ||
31f18b77 | 139 | bool remove = true; |
f6b5b4d7 | 140 | for (auto&& dir : in->get_dirfrags()) { |
31f18b77 FG |
141 | if (!dir->is_auth()) |
142 | continue; | |
143 | ||
f67539c2 TL |
144 | if (export_pin == MDS_RANK_EPHEMERAL_DIST) { |
145 | if (dir->get_frag().bits() < min_frag_bits) { | |
146 | if (!dir->state_test(CDir::STATE_CREATING) && | |
147 | !dir->is_frozen() && !dir->is_freezing()) { | |
148 | queue_split(dir, true); | |
149 | } | |
150 | remove = false; | |
151 | continue; | |
152 | } | |
153 | target = mdcache->hash_into_rank_bucket(in->ino(), dir->get_frag()); | |
154 | } | |
155 | ||
156 | if (target == MDS_RANK_NONE) { | |
31f18b77 FG |
157 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) { |
158 | if (dir->is_frozen() || dir->is_freezing()) { | |
159 | // try again later | |
160 | remove = false; | |
161 | continue; | |
162 | } | |
163 | dout(10) << " clear auxsubtree on " << *dir << dendl; | |
164 | dir->state_clear(CDir::STATE_AUXSUBTREE); | |
165 | mds->mdcache->try_subtree_merge(dir); | |
166 | } | |
f67539c2 | 167 | } else if (target == mds->get_nodeid()) { |
f6b5b4d7 TL |
168 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) { |
169 | ceph_assert(dir->is_subtree_root()); | |
170 | } else if (dir->state_test(CDir::STATE_CREATING) || | |
171 | dir->is_frozen() || dir->is_freezing()) { | |
31f18b77 FG |
172 | // try again later |
173 | remove = false; | |
174 | continue; | |
f6b5b4d7 | 175 | } else if (!dir->is_subtree_root()) { |
31f18b77 FG |
176 | dir->state_set(CDir::STATE_AUXSUBTREE); |
177 | mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid()); | |
178 | dout(10) << " create aux subtree on " << *dir << dendl; | |
f6b5b4d7 | 179 | } else { |
31f18b77 FG |
180 | dout(10) << " set auxsubtree bit on " << *dir << dendl; |
181 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
182 | } | |
183 | } else { | |
f6b5b4d7 TL |
184 | /* Only export a directory if it's non-empty. An empty directory will |
185 | * be sent back by the importer. | |
186 | */ | |
187 | if (dir->get_num_head_items() > 0) { | |
f67539c2 | 188 | mds->mdcache->migrator->export_dir(dir, target); |
f6b5b4d7 | 189 | } |
31f18b77 | 190 | remove = false; |
7c673cae FG |
191 | } |
192 | } | |
31f18b77 FG |
193 | |
194 | if (remove) { | |
195 | in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); | |
196 | q.erase(cur); | |
7c673cae FG |
197 | } |
198 | } | |
199 | ||
f67539c2 | 200 | std::vector<CDir*> authsubs = mdcache->get_auth_subtrees(); |
81eedcae TL |
201 | bool print_auth_subtrees = true; |
202 | ||
203 | if (authsubs.size() > AUTH_TREES_THRESHOLD && | |
204 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
205 | dout(15) << "number of auth trees = " << authsubs.size() << "; not " | |
206 | "printing auth trees" << dendl; | |
207 | print_auth_subtrees = false; | |
208 | } | |
209 | ||
210 | for (auto &cd : authsubs) { | |
7c673cae | 211 | mds_rank_t export_pin = cd->inode->get_export_pin(); |
f67539c2 | 212 | cd->inode->check_pin_policy(export_pin); |
81eedcae | 213 | |
f67539c2 TL |
214 | if (export_pin == MDS_RANK_EPHEMERAL_DIST) { |
215 | export_pin = mdcache->hash_into_rank_bucket(cd->ino(), cd->get_frag()); | |
216 | } else if (export_pin == MDS_RANK_EPHEMERAL_RAND) { | |
217 | export_pin = mdcache->hash_into_rank_bucket(cd->ino()); | |
81eedcae TL |
218 | } |
219 | ||
f67539c2 TL |
220 | if (print_auth_subtrees) |
221 | dout(25) << "auth tree " << *cd << " export_pin=" << export_pin << dendl; | |
222 | ||
223 | if (export_pin >= 0 && export_pin != mds->get_nodeid() && | |
224 | export_pin < mds->mdsmap->get_max_mds()) { | |
225 | mdcache->migrator->export_dir(cd, export_pin); | |
7c673cae FG |
226 | } |
227 | } | |
228 | } | |
229 | ||
230 | void MDBalancer::tick() | |
231 | { | |
11fdf7f2 TL |
232 | static int num_bal_times = g_conf()->mds_bal_max; |
233 | auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); | |
234 | auto bal_max_until = g_conf().get_val<int64_t>("mds_bal_max_until"); | |
235 | time now = clock::now(); | |
7c673cae | 236 | |
11fdf7f2 | 237 | if (g_conf()->mds_bal_export_pin) { |
7c673cae FG |
238 | handle_export_pins(); |
239 | } | |
240 | ||
241 | // sample? | |
11fdf7f2 TL |
242 | if (chrono::duration<double>(now-last_sample).count() > |
243 | g_conf()->mds_bal_sample_interval) { | |
7c673cae FG |
244 | dout(15) << "tick last_sample now " << now << dendl; |
245 | last_sample = now; | |
246 | } | |
247 | ||
11fdf7f2 TL |
248 | // We can use duration_cast below, although the result is an int, |
249 | // because the values from g_conf are also integers. | |
7c673cae | 250 | // balance? |
11fdf7f2 TL |
251 | if (mds->get_nodeid() == 0 |
252 | && mds->is_active() | |
253 | && bal_interval > 0 | |
20effc67 | 254 | && chrono::duration_cast<chrono::seconds>(now - last_heartbeat).count() >= bal_interval |
11fdf7f2 | 255 | && (num_bal_times || (bal_max_until >= 0 && mds->get_uptime().count() > bal_max_until))) { |
7c673cae FG |
256 | last_heartbeat = now; |
257 | send_heartbeat(); | |
258 | num_bal_times--; | |
259 | } | |
81eedcae TL |
260 | |
261 | mds->mdcache->show_subtrees(10, true); | |
7c673cae FG |
262 | } |
263 | ||
264 | ||
265 | ||
266 | ||
267 | class C_Bal_SendHeartbeat : public MDSInternalContext { | |
268 | public: | |
269 | explicit C_Bal_SendHeartbeat(MDSRank *mds_) : MDSInternalContext(mds_) { } | |
270 | void finish(int f) override { | |
271 | mds->balancer->send_heartbeat(); | |
272 | } | |
273 | }; | |
274 | ||
275 | ||
11fdf7f2 | 276 | double mds_load_t::mds_load() const |
7c673cae | 277 | { |
11fdf7f2 | 278 | switch(g_conf()->mds_bal_mode) { |
7c673cae FG |
279 | case 0: |
280 | return | |
281 | .8 * auth.meta_load() + | |
282 | .2 * all.meta_load() + | |
283 | req_rate + | |
284 | 10.0 * queue_len; | |
285 | ||
286 | case 1: | |
287 | return req_rate + 10.0*queue_len; | |
288 | ||
289 | case 2: | |
290 | return cpu_load_avg; | |
291 | ||
292 | } | |
293 | ceph_abort(); | |
294 | return 0; | |
295 | } | |
296 | ||
11fdf7f2 | 297 | mds_load_t MDBalancer::get_load() |
7c673cae | 298 | { |
11fdf7f2 TL |
299 | auto now = clock::now(); |
300 | ||
301 | mds_load_t load{DecayRate()}; /* zero DecayRate! */ | |
7c673cae FG |
302 | |
303 | if (mds->mdcache->get_root()) { | |
9f95a23c | 304 | auto&& ls = mds->mdcache->get_root()->get_dirfrags(); |
11fdf7f2 TL |
305 | for (auto &d : ls) { |
306 | load.auth.add(d->pop_auth_subtree_nested); | |
307 | load.all.add(d->pop_nested); | |
7c673cae FG |
308 | } |
309 | } else { | |
9f95a23c | 310 | dout(20) << "no root, no load" << dendl; |
7c673cae FG |
311 | } |
312 | ||
28e407b8 | 313 | uint64_t num_requests = mds->get_num_requests(); |
522d829b TL |
314 | uint64_t num_traverse = mds->logger->get(l_mds_traverse); |
315 | uint64_t num_traverse_hit = mds->logger->get(l_mds_traverse_hit); | |
91327a77 AA |
316 | |
317 | uint64_t cpu_time = 1; | |
318 | { | |
319 | string stat_path = PROCPREFIX "/proc/self/stat"; | |
320 | ifstream stat_file(stat_path); | |
321 | if (stat_file.is_open()) { | |
322 | vector<string> stat_vec(std::istream_iterator<string>{stat_file}, | |
323 | std::istream_iterator<string>()); | |
324 | if (stat_vec.size() >= 15) { | |
325 | // utime + stime | |
326 | cpu_time = strtoll(stat_vec[13].c_str(), nullptr, 10) + | |
327 | strtoll(stat_vec[14].c_str(), nullptr, 10); | |
328 | } else { | |
329 | derr << "input file '" << stat_path << "' not resolvable" << dendl_impl; | |
330 | } | |
331 | } else { | |
332 | derr << "input file '" << stat_path << "' not found" << dendl_impl; | |
333 | } | |
334 | } | |
335 | ||
336 | load.queue_len = messenger->get_dispatch_queue_len(); | |
337 | ||
338 | bool update_last = true; | |
11fdf7f2 | 339 | if (last_get_load != clock::zero() && |
91327a77 | 340 | now > last_get_load) { |
11fdf7f2 TL |
341 | double el = std::chrono::duration<double>(now-last_get_load).count(); |
342 | if (el >= 1.0) { | |
91327a77 | 343 | if (num_requests > last_num_requests) |
11fdf7f2 | 344 | load.req_rate = (num_requests - last_num_requests) / el; |
91327a77 | 345 | if (cpu_time > last_cpu_time) |
11fdf7f2 | 346 | load.cpu_load_avg = (cpu_time - last_cpu_time) / el; |
522d829b TL |
347 | if (num_traverse > last_num_traverse && num_traverse_hit > last_num_traverse_hit) |
348 | load.cache_hit_rate = (double)(num_traverse_hit - last_num_traverse_hit) / (num_traverse - last_num_traverse); | |
91327a77 AA |
349 | } else { |
350 | auto p = mds_load.find(mds->get_nodeid()); | |
351 | if (p != mds_load.end()) { | |
352 | load.req_rate = p->second.req_rate; | |
353 | load.cpu_load_avg = p->second.cpu_load_avg; | |
522d829b | 354 | load.cache_hit_rate = p->second.cache_hit_rate; |
91327a77 | 355 | } |
522d829b TL |
356 | if (num_requests >= last_num_requests && cpu_time >= last_cpu_time && |
357 | num_traverse >= last_num_traverse && num_traverse_hit >= last_num_traverse_hit) | |
91327a77 | 358 | update_last = false; |
28e407b8 AA |
359 | } |
360 | } | |
28e407b8 | 361 | |
91327a77 AA |
362 | if (update_last) { |
363 | last_num_requests = num_requests; | |
364 | last_cpu_time = cpu_time; | |
365 | last_get_load = now; | |
522d829b TL |
366 | last_num_traverse = num_traverse; |
367 | last_num_traverse_hit = num_traverse_hit; | |
91327a77 | 368 | } |
7c673cae | 369 | |
9f95a23c | 370 | dout(15) << load << dendl; |
7c673cae FG |
371 | return load; |
372 | } | |
373 | ||
374 | /* | |
375 | * Read synchronously from RADOS using a timeout. We cannot do daemon-local | |
376 | * fallbacks (i.e. kick off async read when we are processing the map and | |
377 | * check status when we get here) with the way the mds is structured. | |
378 | */ | |
379 | int MDBalancer::localize_balancer() | |
380 | { | |
381 | /* reset everything */ | |
382 | bool ack = false; | |
383 | int r = 0; | |
384 | bufferlist lua_src; | |
9f95a23c TL |
385 | ceph::mutex lock = ceph::make_mutex("lock"); |
386 | ceph::condition_variable cond; | |
7c673cae FG |
387 | |
388 | /* we assume that balancer is in the metadata pool */ | |
389 | object_t oid = object_t(mds->mdsmap->get_balancer()); | |
b3b6e05e | 390 | object_locator_t oloc(mds->get_metadata_pool()); |
7c673cae | 391 | ceph_tid_t tid = mds->objecter->read(oid, oloc, 0, 0, CEPH_NOSNAP, &lua_src, 0, |
9f95a23c | 392 | new C_SafeCond(lock, cond, &ack, &r)); |
7c673cae FG |
393 | dout(15) << "launched non-blocking read tid=" << tid |
394 | << " oid=" << oid << " oloc=" << oloc << dendl; | |
395 | ||
396 | /* timeout: if we waste half our time waiting for RADOS, then abort! */ | |
9f95a23c TL |
397 | std::cv_status ret_t = [&] { |
398 | auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); | |
399 | std::unique_lock locker{lock}; | |
400 | return cond.wait_for(locker, std::chrono::seconds(bal_interval / 2)); | |
401 | }(); | |
7c673cae FG |
402 | /* success: store the balancer in memory and set the version. */ |
403 | if (!r) { | |
9f95a23c | 404 | if (ret_t == std::cv_status::timeout) { |
f67539c2 TL |
405 | mds->objecter->op_cancel(tid, -CEPHFS_ECANCELED); |
406 | return -CEPHFS_ETIMEDOUT; | |
7c673cae FG |
407 | } |
408 | bal_code.assign(lua_src.to_str()); | |
409 | bal_version.assign(oid.name); | |
9f95a23c | 410 | dout(10) "bal_code=" << bal_code << dendl; |
7c673cae FG |
411 | } |
412 | return r; | |
413 | } | |
414 | ||
415 | void MDBalancer::send_heartbeat() | |
416 | { | |
7c673cae | 417 | if (mds->is_cluster_degraded()) { |
9f95a23c | 418 | dout(10) << "degraded" << dendl; |
7c673cae FG |
419 | return; |
420 | } | |
421 | ||
422 | if (!mds->mdcache->is_open()) { | |
9f95a23c | 423 | dout(10) << "not open" << dendl; |
7c673cae FG |
424 | mds->mdcache->wait_for_open(new C_Bal_SendHeartbeat(mds)); |
425 | return; | |
426 | } | |
427 | ||
94b18763 | 428 | if (mds->get_nodeid() == 0) { |
7c673cae | 429 | beat_epoch++; |
94b18763 FG |
430 | mds_load.clear(); |
431 | } | |
7c673cae FG |
432 | |
433 | // my load | |
11fdf7f2 | 434 | mds_load_t load = get_load(); |
28e407b8 AA |
435 | mds->logger->set(l_mds_load_cent, 100 * load.mds_load()); |
436 | mds->logger->set(l_mds_dispatch_queue_len, load.queue_len); | |
437 | ||
11fdf7f2 TL |
438 | auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(mds->get_nodeid()), std::forward_as_tuple(load)); |
439 | if (!em.second) { | |
440 | em.first->second = load; | |
441 | } | |
7c673cae FG |
442 | |
443 | // import_map -- how much do i import from whom | |
444 | map<mds_rank_t, float> import_map; | |
11fdf7f2 | 445 | for (auto& im : mds->mdcache->get_auth_subtrees()) { |
7c673cae FG |
446 | mds_rank_t from = im->inode->authority().first; |
447 | if (from == mds->get_nodeid()) continue; | |
448 | if (im->get_inode()->is_stray()) continue; | |
11fdf7f2 | 449 | import_map[from] += im->pop_auth_subtree.meta_load(); |
7c673cae FG |
450 | } |
451 | mds_import_map[ mds->get_nodeid() ] = import_map; | |
452 | ||
453 | ||
9f95a23c TL |
454 | dout(3) << " epoch " << beat_epoch << " load " << load << dendl; |
455 | for (const auto& [rank, load] : import_map) { | |
456 | dout(5) << " import_map from " << rank << " -> " << load << dendl; | |
7c673cae FG |
457 | } |
458 | ||
459 | ||
460 | set<mds_rank_t> up; | |
461 | mds->get_mds_map()->get_up_mds_set(up); | |
11fdf7f2 TL |
462 | for (const auto& r : up) { |
463 | if (r == mds->get_nodeid()) | |
7c673cae | 464 | continue; |
9f95a23c | 465 | auto hb = make_message<MHeartbeat>(load, beat_epoch); |
7c673cae | 466 | hb->get_import_map() = import_map; |
11fdf7f2 | 467 | mds->send_message_mds(hb, r); |
7c673cae FG |
468 | } |
469 | } | |
470 | ||
9f95a23c | 471 | void MDBalancer::handle_heartbeat(const cref_t<MHeartbeat> &m) |
7c673cae | 472 | { |
7c673cae FG |
473 | mds_rank_t who = mds_rank_t(m->get_source().num()); |
474 | dout(25) << "=== got heartbeat " << m->get_beat() << " from " << m->get_source().num() << " " << m->get_load() << dendl; | |
475 | ||
476 | if (!mds->is_active()) | |
11fdf7f2 | 477 | return; |
7c673cae FG |
478 | |
479 | if (!mds->mdcache->is_open()) { | |
480 | dout(10) << "opening root on handle_heartbeat" << dendl; | |
481 | mds->mdcache->wait_for_open(new C_MDS_RetryMessage(mds, m)); | |
482 | return; | |
483 | } | |
484 | ||
485 | if (mds->is_cluster_degraded()) { | |
486 | dout(10) << " degraded, ignoring" << dendl; | |
11fdf7f2 | 487 | return; |
7c673cae FG |
488 | } |
489 | ||
94b18763 FG |
490 | if (mds->get_nodeid() != 0 && m->get_beat() > beat_epoch) { |
491 | dout(10) << "receive next epoch " << m->get_beat() << " from mds." << who << " before mds0" << dendl; | |
492 | ||
493 | beat_epoch = m->get_beat(); | |
494 | // clear the mds load info whose epoch is less than beat_epoch | |
495 | mds_load.clear(); | |
496 | } | |
497 | ||
7c673cae | 498 | if (who == 0) { |
94b18763 FG |
499 | dout(20) << " from mds0, new epoch " << m->get_beat() << dendl; |
500 | if (beat_epoch != m->get_beat()) { | |
28e407b8 | 501 | beat_epoch = m->get_beat(); |
94b18763 FG |
502 | mds_load.clear(); |
503 | } | |
28e407b8 | 504 | |
7c673cae FG |
505 | send_heartbeat(); |
506 | ||
507 | mds->mdcache->show_subtrees(); | |
28e407b8 AA |
508 | } else if (mds->get_nodeid() == 0) { |
509 | if (beat_epoch != m->get_beat()) { | |
510 | dout(10) << " old heartbeat epoch, ignoring" << dendl; | |
11fdf7f2 | 511 | return; |
7c673cae FG |
512 | } |
513 | } | |
28e407b8 | 514 | |
11fdf7f2 TL |
515 | { |
516 | auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(who), std::forward_as_tuple(m->get_load())); | |
517 | if (!em.second) { | |
518 | em.first->second = m->get_load(); | |
519 | } | |
520 | } | |
28e407b8 | 521 | mds_import_map[who] = m->get_import_map(); |
7c673cae | 522 | |
1e59de90 TL |
523 | mds->mdsmap->update_num_mdss_in_rank_mask_bitset(); |
524 | ||
525 | if (mds->mdsmap->get_num_mdss_in_rank_mask_bitset() > 0) | |
7c673cae FG |
526 | { |
527 | unsigned cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
528 | if (mds_load.size() == cluster_size) { | |
529 | // let's go! | |
530 | //export_empties(); // no! | |
531 | ||
532 | /* avoid spamming ceph -w if user does not turn mantle on */ | |
533 | if (mds->mdsmap->get_balancer() != "") { | |
534 | int r = mantle_prep_rebalance(); | |
11fdf7f2 | 535 | if (!r) return; |
7c673cae FG |
536 | mds->clog->warn() << "using old balancer; mantle failed for " |
537 | << "balancer=" << mds->mdsmap->get_balancer() | |
538 | << " : " << cpp_strerror(r); | |
539 | } | |
540 | prep_rebalance(m->get_beat()); | |
541 | } | |
542 | } | |
7c673cae FG |
543 | } |
544 | ||
7c673cae FG |
545 | double MDBalancer::try_match(balance_state_t& state, mds_rank_t ex, double& maxex, |
546 | mds_rank_t im, double& maxim) | |
547 | { | |
548 | if (maxex <= 0 || maxim <= 0) return 0.0; | |
549 | ||
11fdf7f2 | 550 | double howmuch = std::min(maxex, maxim); |
7c673cae FG |
551 | |
552 | dout(5) << " - mds." << ex << " exports " << howmuch << " to mds." << im << dendl; | |
553 | ||
554 | if (ex == mds->get_nodeid()) | |
555 | state.targets[im] += howmuch; | |
556 | ||
557 | state.exported[ex] += howmuch; | |
558 | state.imported[im] += howmuch; | |
559 | ||
560 | maxex -= howmuch; | |
561 | maxim -= howmuch; | |
562 | ||
563 | return howmuch; | |
564 | } | |
565 | ||
566 | void MDBalancer::queue_split(const CDir *dir, bool fast) | |
567 | { | |
568 | dout(10) << __func__ << " enqueuing " << *dir | |
569 | << " (fast=" << fast << ")" << dendl; | |
570 | ||
f67539c2 | 571 | const dirfrag_t df = dir->dirfrag(); |
7c673cae | 572 | |
f67539c2 TL |
573 | auto callback = [this, df](int r) { |
574 | if (split_pending.erase(df) == 0) { | |
7c673cae FG |
575 | // Someone beat me to it. This can happen in the fast splitting |
576 | // path, because we spawn two contexts, one with mds->timer and | |
577 | // one with mds->queue_waiter. The loser can safely just drop | |
578 | // out. | |
579 | return; | |
580 | } | |
581 | ||
f67539c2 TL |
582 | auto mdcache = mds->mdcache; |
583 | ||
584 | CDir *dir = mdcache->get_dirfrag(df); | |
585 | if (!dir) { | |
586 | dout(10) << "drop split on " << df << " because not in cache" << dendl; | |
7c673cae FG |
587 | return; |
588 | } | |
f67539c2 TL |
589 | if (!dir->is_auth()) { |
590 | dout(10) << "drop split on " << df << " because non-auth" << dendl; | |
7c673cae FG |
591 | return; |
592 | } | |
593 | ||
594 | // Pass on to MDCache: note that the split might still not | |
595 | // happen if the checks in MDCache::can_fragment fail. | |
f67539c2 TL |
596 | dout(10) << __func__ << " splitting " << *dir << dendl; |
597 | int bits = g_conf()->mds_bal_split_bits; | |
598 | if (dir->inode->is_ephemeral_dist()) { | |
599 | unsigned min_frag_bits = mdcache->get_ephemeral_dist_frag_bits(); | |
600 | if (df.frag.bits() + bits < min_frag_bits) | |
601 | bits = min_frag_bits - df.frag.bits(); | |
602 | } | |
603 | mdcache->split_dir(dir, bits); | |
7c673cae FG |
604 | }; |
605 | ||
f67539c2 TL |
606 | auto ret = split_pending.insert(df); |
607 | bool is_new = ret.second; | |
7c673cae FG |
608 | |
609 | if (fast) { | |
610 | // Do the split ASAP: enqueue it in the MDSRank waiters which are | |
611 | // run at the end of dispatching the current request | |
612 | mds->queue_waiter(new MDSInternalContextWrapper(mds, | |
9f95a23c | 613 | new LambdaContext(std::move(callback)))); |
7c673cae FG |
614 | } else if (is_new) { |
615 | // Set a timer to really do the split: we don't do it immediately | |
616 | // so that bursts of ops on a directory have a chance to go through | |
617 | // before we freeze it. | |
91327a77 | 618 | mds->timer.add_event_after(bal_fragment_interval, |
9f95a23c | 619 | new LambdaContext(std::move(callback))); |
7c673cae FG |
620 | } |
621 | } | |
622 | ||
623 | void MDBalancer::queue_merge(CDir *dir) | |
624 | { | |
625 | const auto frag = dir->dirfrag(); | |
626 | auto callback = [this, frag](int r) { | |
11fdf7f2 | 627 | ceph_assert(frag.frag != frag_t()); |
7c673cae FG |
628 | |
629 | // frag must be in this set because only one context is in flight | |
630 | // for a given frag at a time (because merge_pending is checked before | |
631 | // starting one), and this context is the only one that erases it. | |
632 | merge_pending.erase(frag); | |
633 | ||
f67539c2 TL |
634 | auto mdcache = mds->mdcache; |
635 | CDir *dir = mdcache->get_dirfrag(frag); | |
7c673cae FG |
636 | if (!dir) { |
637 | dout(10) << "drop merge on " << frag << " because not in cache" << dendl; | |
638 | return; | |
639 | } | |
11fdf7f2 | 640 | ceph_assert(dir->dirfrag() == frag); |
7c673cae FG |
641 | |
642 | if(!dir->is_auth()) { | |
643 | dout(10) << "drop merge on " << *dir << " because lost auth" << dendl; | |
644 | return; | |
645 | } | |
646 | ||
647 | dout(10) << "merging " << *dir << dendl; | |
648 | ||
649 | CInode *diri = dir->get_inode(); | |
650 | ||
f67539c2 TL |
651 | unsigned min_frag_bits = 0; |
652 | if (diri->is_ephemeral_dist()) | |
653 | min_frag_bits = mdcache->get_ephemeral_dist_frag_bits(); | |
654 | ||
7c673cae | 655 | frag_t fg = dir->get_frag(); |
f67539c2 | 656 | while (fg.bits() > min_frag_bits) { |
7c673cae | 657 | frag_t sibfg = fg.get_sibling(); |
9f95a23c | 658 | auto&& [complete, sibs] = diri->get_dirfrags_under(sibfg); |
7c673cae FG |
659 | if (!complete) { |
660 | dout(10) << " not all sibs under " << sibfg << " in cache (have " << sibs << ")" << dendl; | |
661 | break; | |
662 | } | |
663 | bool all = true; | |
9f95a23c | 664 | for (auto& sib : sibs) { |
7c673cae FG |
665 | if (!sib->is_auth() || !sib->should_merge()) { |
666 | all = false; | |
667 | break; | |
668 | } | |
669 | } | |
670 | if (!all) { | |
671 | dout(10) << " not all sibs under " << sibfg << " " << sibs << " should_merge" << dendl; | |
672 | break; | |
673 | } | |
674 | dout(10) << " all sibs under " << sibfg << " " << sibs << " should merge" << dendl; | |
675 | fg = fg.parent(); | |
676 | } | |
677 | ||
678 | if (fg != dir->get_frag()) | |
f67539c2 | 679 | mdcache->merge_dir(diri, fg); |
7c673cae FG |
680 | }; |
681 | ||
682 | if (merge_pending.count(frag) == 0) { | |
9f95a23c | 683 | dout(20) << " enqueued dir " << *dir << dendl; |
7c673cae | 684 | merge_pending.insert(frag); |
91327a77 | 685 | mds->timer.add_event_after(bal_fragment_interval, |
9f95a23c | 686 | new LambdaContext(std::move(callback))); |
7c673cae | 687 | } else { |
9f95a23c | 688 | dout(20) << " dir already in queue " << *dir << dendl; |
7c673cae FG |
689 | } |
690 | } | |
691 | ||
692 | void MDBalancer::prep_rebalance(int beat) | |
693 | { | |
694 | balance_state_t state; | |
695 | ||
11fdf7f2 | 696 | if (g_conf()->mds_thrash_exports) { |
7c673cae FG |
697 | //we're going to randomly export to all the mds in the cluster |
698 | set<mds_rank_t> up_mds; | |
699 | mds->get_mds_map()->get_up_mds_set(up_mds); | |
700 | for (const auto &rank : up_mds) { | |
701 | state.targets[rank] = 0.0; | |
702 | } | |
703 | } else { | |
704 | int cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
705 | mds_rank_t whoami = mds->get_nodeid(); | |
11fdf7f2 | 706 | rebalance_time = clock::now(); |
7c673cae | 707 | |
9f95a23c | 708 | dout(7) << "cluster loads are" << dendl; |
7c673cae FG |
709 | |
710 | mds->mdcache->migrator->clear_export_queue(); | |
711 | ||
712 | // rescale! turn my mds_load back into meta_load units | |
713 | double load_fac = 1.0; | |
714 | map<mds_rank_t, mds_load_t>::iterator m = mds_load.find(whoami); | |
715 | if ((m != mds_load.end()) && (m->second.mds_load() > 0)) { | |
11fdf7f2 | 716 | double metald = m->second.auth.meta_load(); |
7c673cae FG |
717 | double mdsld = m->second.mds_load(); |
718 | load_fac = metald / mdsld; | |
719 | dout(7) << " load_fac is " << load_fac | |
720 | << " <- " << m->second.auth << " " << metald | |
721 | << " / " << mdsld | |
722 | << dendl; | |
723 | } | |
724 | ||
28e407b8 AA |
725 | mds_meta_load.clear(); |
726 | ||
7c673cae FG |
727 | double total_load = 0.0; |
728 | multimap<double,mds_rank_t> load_map; | |
729 | for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { | |
28e407b8 | 730 | mds_load_t& load = mds_load.at(i); |
7c673cae FG |
731 | |
732 | double l = load.mds_load() * load_fac; | |
733 | mds_meta_load[i] = l; | |
734 | ||
735 | if (whoami == 0) | |
9f95a23c | 736 | dout(7) << " mds." << i |
7c673cae FG |
737 | << " " << load |
738 | << " = " << load.mds_load() | |
739 | << " ~ " << l << dendl; | |
740 | ||
741 | if (whoami == i) my_load = l; | |
742 | total_load += l; | |
743 | ||
744 | load_map.insert(pair<double,mds_rank_t>( l, i )); | |
745 | } | |
746 | ||
747 | // target load | |
1e59de90 | 748 | target_load = total_load / (double)mds->mdsmap->get_num_mdss_in_rank_mask_bitset(); |
9f95a23c | 749 | dout(7) << "my load " << my_load |
7c673cae FG |
750 | << " target " << target_load |
751 | << " total " << total_load | |
752 | << dendl; | |
753 | ||
754 | // under or over? | |
9f95a23c | 755 | for (const auto& [load, rank] : load_map) { |
1e59de90 TL |
756 | if (test_rank_mask(rank) && |
757 | load < target_load * (1.0 + g_conf()->mds_bal_min_rebalance)) { | |
9f95a23c TL |
758 | dout(7) << " mds." << rank << " is underloaded or barely overloaded." << dendl; |
759 | mds_last_epoch_under_map[rank] = beat_epoch; | |
28e407b8 AA |
760 | } |
761 | } | |
762 | ||
763 | int last_epoch_under = mds_last_epoch_under_map[whoami]; | |
764 | if (last_epoch_under == beat_epoch) { | |
9f95a23c | 765 | dout(7) << " i am underloaded or barely overloaded, doing nothing." << dendl; |
7c673cae FG |
766 | return; |
767 | } | |
7c673cae FG |
768 | // am i over long enough? |
769 | if (last_epoch_under && beat_epoch - last_epoch_under < 2) { | |
9f95a23c | 770 | dout(7) << " i am overloaded, but only for " << (beat_epoch - last_epoch_under) << " epochs" << dendl; |
7c673cae FG |
771 | return; |
772 | } | |
773 | ||
9f95a23c | 774 | dout(7) << " i am sufficiently overloaded" << dendl; |
7c673cae FG |
775 | |
776 | ||
777 | // first separate exporters and importers | |
778 | multimap<double,mds_rank_t> importers; | |
779 | multimap<double,mds_rank_t> exporters; | |
780 | set<mds_rank_t> importer_set; | |
781 | set<mds_rank_t> exporter_set; | |
782 | ||
783 | for (multimap<double,mds_rank_t>::iterator it = load_map.begin(); | |
784 | it != load_map.end(); | |
785 | ++it) { | |
1e59de90 | 786 | if (it->first < target_load && test_rank_mask(it->second)) { |
7c673cae FG |
787 | dout(15) << " mds." << it->second << " is importer" << dendl; |
788 | importers.insert(pair<double,mds_rank_t>(it->first,it->second)); | |
789 | importer_set.insert(it->second); | |
790 | } else { | |
28e407b8 AA |
791 | int mds_last_epoch_under = mds_last_epoch_under_map[it->second]; |
792 | if (!(mds_last_epoch_under && beat_epoch - mds_last_epoch_under < 2)) { | |
793 | dout(15) << " mds." << it->second << " is exporter" << dendl; | |
794 | exporters.insert(pair<double,mds_rank_t>(it->first,it->second)); | |
795 | exporter_set.insert(it->second); | |
796 | } | |
7c673cae FG |
797 | } |
798 | } | |
799 | ||
800 | ||
801 | // determine load transfer mapping | |
802 | ||
803 | if (true) { | |
804 | // analyze import_map; do any matches i can | |
805 | ||
806 | dout(15) << " matching exporters to import sources" << dendl; | |
807 | ||
808 | // big -> small exporters | |
809 | for (multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); | |
810 | ex != exporters.rend(); | |
811 | ++ex) { | |
1e59de90 TL |
812 | double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0; |
813 | double maxex = get_maxex(state, ex->second, ex_target_load); | |
7c673cae FG |
814 | if (maxex <= .001) continue; |
815 | ||
816 | // check importers. for now, just in arbitrary order (no intelligent matching). | |
817 | for (map<mds_rank_t, float>::iterator im = mds_import_map[ex->second].begin(); | |
818 | im != mds_import_map[ex->second].end(); | |
819 | ++im) { | |
1e59de90 | 820 | double maxim = get_maxim(state, im->first, target_load); |
7c673cae FG |
821 | if (maxim <= .001) continue; |
822 | try_match(state, ex->second, maxex, im->first, maxim); | |
823 | if (maxex <= .001) break; | |
824 | } | |
825 | } | |
826 | } | |
827 | ||
828 | // old way | |
829 | if (beat % 2 == 1) { | |
830 | dout(15) << " matching big exporters to big importers" << dendl; | |
831 | // big exporters to big importers | |
832 | multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); | |
833 | multimap<double,mds_rank_t>::iterator im = importers.begin(); | |
834 | while (ex != exporters.rend() && | |
835 | im != importers.end()) { | |
1e59de90 TL |
836 | double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0; |
837 | double maxex = get_maxex(state, ex->second, ex_target_load); | |
838 | double maxim = get_maxim(state, im->second, target_load); | |
7c673cae FG |
839 | if (maxex < .001 || maxim < .001) break; |
840 | try_match(state, ex->second, maxex, im->second, maxim); | |
841 | if (maxex <= .001) ++ex; | |
842 | if (maxim <= .001) ++im; | |
843 | } | |
844 | } else { // new way | |
845 | dout(15) << " matching small exporters to big importers" << dendl; | |
846 | // small exporters to big importers | |
847 | multimap<double,mds_rank_t>::iterator ex = exporters.begin(); | |
848 | multimap<double,mds_rank_t>::iterator im = importers.begin(); | |
849 | while (ex != exporters.end() && | |
850 | im != importers.end()) { | |
1e59de90 TL |
851 | double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0; |
852 | double maxex = get_maxex(state, ex->second, ex_target_load); | |
853 | double maxim = get_maxim(state, im->second, target_load); | |
7c673cae FG |
854 | if (maxex < .001 || maxim < .001) break; |
855 | try_match(state, ex->second, maxex, im->second, maxim); | |
856 | if (maxex <= .001) ++ex; | |
857 | if (maxim <= .001) ++im; | |
858 | } | |
859 | } | |
860 | } | |
861 | try_rebalance(state); | |
862 | } | |
863 | ||
7c673cae FG |
864 | int MDBalancer::mantle_prep_rebalance() |
865 | { | |
866 | balance_state_t state; | |
867 | ||
868 | /* refresh balancer if it has changed */ | |
869 | if (bal_version != mds->mdsmap->get_balancer()) { | |
870 | bal_version.assign(""); | |
871 | int r = localize_balancer(); | |
872 | if (r) return r; | |
873 | ||
874 | /* only spam the cluster log from 1 mds on version changes */ | |
875 | if (mds->get_nodeid() == 0) | |
876 | mds->clog->info() << "mantle balancer version changed: " << bal_version; | |
877 | } | |
878 | ||
879 | /* prepare for balancing */ | |
880 | int cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
11fdf7f2 | 881 | rebalance_time = clock::now(); |
7c673cae FG |
882 | mds->mdcache->migrator->clear_export_queue(); |
883 | ||
884 | /* fill in the metrics for each mds by grabbing load struct */ | |
885 | vector < map<string, double> > metrics (cluster_size); | |
28e407b8 AA |
886 | for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { |
887 | mds_load_t& load = mds_load.at(i); | |
7c673cae FG |
888 | |
889 | metrics[i] = {{"auth.meta_load", load.auth.meta_load()}, | |
890 | {"all.meta_load", load.all.meta_load()}, | |
891 | {"req_rate", load.req_rate}, | |
892 | {"queue_len", load.queue_len}, | |
893 | {"cpu_load_avg", load.cpu_load_avg}}; | |
894 | } | |
895 | ||
896 | /* execute the balancer */ | |
897 | Mantle mantle; | |
898 | int ret = mantle.balance(bal_code, mds->get_nodeid(), metrics, state.targets); | |
9f95a23c | 899 | dout(7) << " mantle decided that new targets=" << state.targets << dendl; |
7c673cae FG |
900 | |
901 | /* mantle doesn't know about cluster size, so check target len here */ | |
902 | if ((int) state.targets.size() != cluster_size) | |
f67539c2 | 903 | return -CEPHFS_EINVAL; |
7c673cae FG |
904 | else if (ret) |
905 | return ret; | |
906 | ||
907 | try_rebalance(state); | |
908 | return 0; | |
909 | } | |
910 | ||
911 | ||
912 | ||
913 | void MDBalancer::try_rebalance(balance_state_t& state) | |
914 | { | |
11fdf7f2 | 915 | if (g_conf()->mds_thrash_exports) { |
7c673cae FG |
916 | dout(5) << "mds_thrash is on; not performing standard rebalance operation!" |
917 | << dendl; | |
918 | return; | |
919 | } | |
920 | ||
921 | // make a sorted list of my imports | |
28e407b8 AA |
922 | multimap<double, CDir*> import_pop_map; |
923 | multimap<mds_rank_t, pair<CDir*, double> > import_from_map; | |
7c673cae | 924 | |
11fdf7f2 | 925 | for (auto& dir : mds->mdcache->get_fullauth_subtrees()) { |
28e407b8 AA |
926 | CInode *diri = dir->get_inode(); |
927 | if (diri->is_mdsdir()) | |
928 | continue; | |
929 | if (diri->get_export_pin(false) != MDS_RANK_NONE) | |
930 | continue; | |
931 | if (dir->is_freezing() || dir->is_frozen()) | |
932 | continue; // export pbly already in progress | |
7c673cae | 933 | |
28e407b8 | 934 | mds_rank_t from = diri->authority().first; |
11fdf7f2 TL |
935 | double pop = dir->pop_auth_subtree.meta_load(); |
936 | if (g_conf()->mds_bal_idle_threshold > 0 && | |
937 | pop < g_conf()->mds_bal_idle_threshold && | |
28e407b8 AA |
938 | diri != mds->mdcache->get_root() && |
939 | from != mds->get_nodeid()) { | |
940 | dout(5) << " exporting idle (" << pop << ") import " << *dir | |
941 | << " back to mds." << from << dendl; | |
942 | mds->mdcache->migrator->export_dir_nicely(dir, from); | |
7c673cae FG |
943 | continue; |
944 | } | |
945 | ||
28e407b8 AA |
946 | dout(15) << " map: i imported " << *dir << " from " << from << dendl; |
947 | import_pop_map.insert(make_pair(pop, dir)); | |
948 | import_from_map.insert(make_pair(from, make_pair(dir, pop))); | |
7c673cae FG |
949 | } |
950 | ||
7c673cae | 951 | // do my exports! |
28e407b8 | 952 | map<mds_rank_t, double> export_pop_map; |
7c673cae FG |
953 | |
954 | for (auto &it : state.targets) { | |
955 | mds_rank_t target = it.first; | |
956 | double amount = it.second; | |
957 | ||
1e59de90 | 958 | if (amount < MIN_OFFLOAD) { |
28e407b8 | 959 | continue; |
1e59de90 TL |
960 | } |
961 | if (amount * 10 * state.targets.size() < target_load) { | |
91327a77 | 962 | continue; |
1e59de90 | 963 | } |
7c673cae FG |
964 | |
965 | dout(5) << "want to send " << amount << " to mds." << target | |
966 | //<< " .. " << (*it).second << " * " << load_fac | |
967 | << " -> " << amount | |
968 | << dendl;//" .. fudge is " << fudge << dendl; | |
7c673cae | 969 | |
28e407b8 | 970 | double& have = export_pop_map[target]; |
7c673cae FG |
971 | |
972 | mds->mdcache->show_subtrees(); | |
973 | ||
974 | // search imports from target | |
975 | if (import_from_map.count(target)) { | |
9f95a23c | 976 | dout(7) << " aha, looking through imports from target mds." << target << dendl; |
28e407b8 AA |
977 | for (auto p = import_from_map.equal_range(target); |
978 | p.first != p.second; ) { | |
979 | CDir *dir = p.first->second.first; | |
980 | double pop = p.first->second.second; | |
9f95a23c | 981 | dout(7) << "considering " << *dir << " from " << (*p.first).first << dendl; |
28e407b8 | 982 | auto plast = p.first++; |
7c673cae | 983 | |
28e407b8 | 984 | if (dir->inode->is_base()) |
7c673cae | 985 | continue; |
11fdf7f2 | 986 | ceph_assert(dir->inode->authority().first == target); // cuz that's how i put it in the map, dummy |
7c673cae FG |
987 | |
988 | if (pop <= amount-have) { | |
9f95a23c | 989 | dout(7) << "reexporting " << *dir << " pop " << pop |
7c673cae FG |
990 | << " back to mds." << target << dendl; |
991 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
992 | have += pop; | |
993 | import_from_map.erase(plast); | |
28e407b8 AA |
994 | for (auto q = import_pop_map.equal_range(pop); |
995 | q.first != q.second; ) { | |
996 | if (q.first->second == dir) { | |
997 | import_pop_map.erase(q.first); | |
998 | break; | |
999 | } | |
1000 | q.first++; | |
1001 | } | |
7c673cae | 1002 | } else { |
9f95a23c | 1003 | dout(7) << "can't reexport " << *dir << ", too big " << pop << dendl; |
7c673cae | 1004 | } |
28e407b8 AA |
1005 | if (amount-have < MIN_OFFLOAD) |
1006 | break; | |
7c673cae FG |
1007 | } |
1008 | } | |
28e407b8 AA |
1009 | } |
1010 | ||
1011 | // any other imports | |
1012 | for (auto &it : state.targets) { | |
1013 | mds_rank_t target = it.first; | |
1014 | double amount = it.second; | |
1015 | ||
1016 | if (!export_pop_map.count(target)) | |
1017 | continue; | |
1018 | double& have = export_pop_map[target]; | |
1019 | if (amount-have < MIN_OFFLOAD) | |
7c673cae | 1020 | continue; |
7c673cae | 1021 | |
28e407b8 AA |
1022 | for (auto p = import_pop_map.begin(); |
1023 | p != import_pop_map.end(); ) { | |
1024 | CDir *dir = p->second; | |
1025 | if (dir->inode->is_base()) { | |
1026 | ++p; | |
1027 | continue; | |
1028 | } | |
7c673cae | 1029 | |
28e407b8 AA |
1030 | double pop = p->first; |
1031 | if (pop <= amount-have && pop > MIN_REEXPORT) { | |
9f95a23c | 1032 | dout(5) << "reexporting " << *dir << " pop " << pop |
28e407b8 AA |
1033 | << " to mds." << target << dendl; |
1034 | have += pop; | |
1035 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
1036 | import_pop_map.erase(p++); | |
1037 | } else { | |
1038 | ++p; | |
7c673cae | 1039 | } |
28e407b8 AA |
1040 | if (amount-have < MIN_OFFLOAD) |
1041 | break; | |
7c673cae | 1042 | } |
28e407b8 | 1043 | } |
7c673cae | 1044 | |
28e407b8 AA |
1045 | set<CDir*> already_exporting; |
1046 | ||
1047 | for (auto &it : state.targets) { | |
1048 | mds_rank_t target = it.first; | |
1049 | double amount = it.second; | |
7c673cae | 1050 | |
28e407b8 AA |
1051 | if (!export_pop_map.count(target)) |
1052 | continue; | |
1053 | double& have = export_pop_map[target]; | |
1054 | if (amount-have < MIN_OFFLOAD) | |
1055 | continue; | |
1056 | ||
1057 | // okay, search for fragments of my workload | |
9f95a23c | 1058 | std::vector<CDir*> exports; |
7c673cae | 1059 | |
28e407b8 AA |
1060 | for (auto p = import_pop_map.rbegin(); |
1061 | p != import_pop_map.rend(); | |
1062 | ++p) { | |
1063 | CDir *dir = p->second; | |
9f95a23c | 1064 | find_exports(dir, amount, &exports, have, already_exporting); |
28e407b8 | 1065 | if (amount-have < MIN_OFFLOAD) |
7c673cae FG |
1066 | break; |
1067 | } | |
1068 | //fudge = amount - have; | |
1069 | ||
9f95a23c | 1070 | for (const auto& dir : exports) { |
28e407b8 | 1071 | dout(5) << " - exporting " << dir->pop_auth_subtree |
11fdf7f2 | 1072 | << " " << dir->pop_auth_subtree.meta_load() |
28e407b8 AA |
1073 | << " to mds." << target << " " << *dir << dendl; |
1074 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
7c673cae FG |
1075 | } |
1076 | } | |
1077 | ||
9f95a23c | 1078 | dout(7) << "done" << dendl; |
7c673cae FG |
1079 | mds->mdcache->show_subtrees(); |
1080 | } | |
1081 | ||
7c673cae FG |
1082 | void MDBalancer::find_exports(CDir *dir, |
1083 | double amount, | |
9f95a23c | 1084 | std::vector<CDir*>* exports, |
7c673cae FG |
1085 | double& have, |
1086 | set<CDir*>& already_exporting) | |
1087 | { | |
11fdf7f2 TL |
1088 | auto now = clock::now(); |
1089 | auto duration = std::chrono::duration<double>(now-rebalance_time).count(); | |
1090 | if (duration > 0.1) { | |
28e407b8 AA |
1091 | derr << " balancer runs too long" << dendl_impl; |
1092 | have = amount; | |
1093 | return; | |
1094 | } | |
1095 | ||
11fdf7f2 | 1096 | ceph_assert(dir->is_auth()); |
28e407b8 | 1097 | |
7c673cae | 1098 | double need = amount - have; |
11fdf7f2 | 1099 | if (need < amount * g_conf()->mds_bal_min_start) |
7c673cae | 1100 | return; // good enough! |
28e407b8 | 1101 | |
11fdf7f2 TL |
1102 | double needmax = need * g_conf()->mds_bal_need_max; |
1103 | double needmin = need * g_conf()->mds_bal_need_min; | |
1104 | double midchunk = need * g_conf()->mds_bal_midchunk; | |
1105 | double minchunk = need * g_conf()->mds_bal_minchunk; | |
7c673cae | 1106 | |
9f95a23c | 1107 | std::vector<CDir*> bigger_rep, bigger_unrep; |
7c673cae FG |
1108 | multimap<double, CDir*> smaller; |
1109 | ||
11fdf7f2 | 1110 | double dir_pop = dir->pop_auth_subtree.meta_load(); |
9f95a23c | 1111 | dout(7) << "in " << dir_pop << " " << *dir << " need " << need << " (" << needmin << " - " << needmax << ")" << dendl; |
7c673cae FG |
1112 | |
1113 | double subdir_sum = 0; | |
28e407b8 AA |
1114 | for (elist<CInode*>::iterator it = dir->pop_lru_subdirs.begin_use_current(); |
1115 | !it.end(); ) { | |
1116 | CInode *in = *it; | |
1117 | ++it; | |
1118 | ||
11fdf7f2 TL |
1119 | ceph_assert(in->is_dir()); |
1120 | ceph_assert(in->get_parent_dir() == dir); | |
7c673cae | 1121 | |
9f95a23c | 1122 | auto&& dfls = in->get_nested_dirfrags(); |
28e407b8 AA |
1123 | |
1124 | size_t num_idle_frags = 0; | |
9f95a23c | 1125 | for (const auto& subdir : dfls) { |
28e407b8 AA |
1126 | if (already_exporting.count(subdir)) |
1127 | continue; | |
7c673cae | 1128 | |
28e407b8 AA |
1129 | // we know all ancestor dirfrags up to subtree root are not freezing or frozen. |
1130 | // It's more efficient to use CDir::is_{freezing,frozen}_tree_root() | |
1131 | if (subdir->is_frozen_dir() || subdir->is_frozen_tree_root() || | |
1132 | subdir->is_freezing_dir() || subdir->is_freezing_tree_root()) | |
1133 | continue; // can't export this right now! | |
7c673cae FG |
1134 | |
1135 | // how popular? | |
11fdf7f2 | 1136 | double pop = subdir->pop_auth_subtree.meta_load(); |
7c673cae FG |
1137 | subdir_sum += pop; |
1138 | dout(15) << " subdir pop " << pop << " " << *subdir << dendl; | |
1139 | ||
28e407b8 AA |
1140 | if (pop < minchunk) { |
1141 | num_idle_frags++; | |
1142 | continue; | |
1143 | } | |
7c673cae FG |
1144 | |
1145 | // lucky find? | |
1146 | if (pop > needmin && pop < needmax) { | |
9f95a23c | 1147 | exports->push_back(subdir); |
7c673cae FG |
1148 | already_exporting.insert(subdir); |
1149 | have += pop; | |
1150 | return; | |
1151 | } | |
1152 | ||
1153 | if (pop > need) { | |
1154 | if (subdir->is_rep()) | |
1155 | bigger_rep.push_back(subdir); | |
1156 | else | |
1157 | bigger_unrep.push_back(subdir); | |
1158 | } else | |
1159 | smaller.insert(pair<double,CDir*>(pop, subdir)); | |
1160 | } | |
28e407b8 AA |
1161 | if (dfls.size() == num_idle_frags) |
1162 | in->item_pop_lru.remove_myself(); | |
7c673cae FG |
1163 | } |
1164 | dout(15) << " sum " << subdir_sum << " / " << dir_pop << dendl; | |
1165 | ||
1166 | // grab some sufficiently big small items | |
1167 | multimap<double,CDir*>::reverse_iterator it; | |
1168 | for (it = smaller.rbegin(); | |
1169 | it != smaller.rend(); | |
1170 | ++it) { | |
1171 | ||
1172 | if ((*it).first < midchunk) | |
1173 | break; // try later | |
1174 | ||
1175 | dout(7) << " taking smaller " << *(*it).second << dendl; | |
1176 | ||
9f95a23c | 1177 | exports->push_back((*it).second); |
7c673cae FG |
1178 | already_exporting.insert((*it).second); |
1179 | have += (*it).first; | |
1180 | if (have > needmin) | |
1181 | return; | |
1182 | } | |
1183 | ||
1184 | // apprently not enough; drill deeper into the hierarchy (if non-replicated) | |
9f95a23c TL |
1185 | for (const auto& dir : bigger_unrep) { |
1186 | dout(15) << " descending into " << *dir << dendl; | |
1187 | find_exports(dir, amount, exports, have, already_exporting); | |
7c673cae FG |
1188 | if (have > needmin) |
1189 | return; | |
1190 | } | |
1191 | ||
1192 | // ok fine, use smaller bits | |
1193 | for (; | |
1194 | it != smaller.rend(); | |
1195 | ++it) { | |
1196 | dout(7) << " taking (much) smaller " << it->first << " " << *(*it).second << dendl; | |
1197 | ||
9f95a23c | 1198 | exports->push_back((*it).second); |
7c673cae FG |
1199 | already_exporting.insert((*it).second); |
1200 | have += (*it).first; | |
1201 | if (have > needmin) | |
1202 | return; | |
1203 | } | |
1204 | ||
1205 | // ok fine, drill into replicated dirs | |
9f95a23c TL |
1206 | for (const auto& dir : bigger_rep) { |
1207 | dout(7) << " descending into replicated " << *dir << dendl; | |
1208 | find_exports(dir, amount, exports, have, already_exporting); | |
7c673cae FG |
1209 | if (have > needmin) |
1210 | return; | |
1211 | } | |
7c673cae FG |
1212 | } |
1213 | ||
1e59de90 | 1214 | void MDBalancer::hit_inode(CInode *in, int type) |
7c673cae FG |
1215 | { |
1216 | // hit inode | |
11fdf7f2 | 1217 | in->pop.get(type).hit(); |
7c673cae FG |
1218 | |
1219 | if (in->get_parent_dn()) | |
1e59de90 | 1220 | hit_dir(in->get_parent_dn()->get_dir(), type); |
7c673cae FG |
1221 | } |
1222 | ||
1223 | void MDBalancer::maybe_fragment(CDir *dir, bool hot) | |
1224 | { | |
1225 | // split/merge | |
11fdf7f2 | 1226 | if (bal_fragment_dirs && bal_fragment_interval > 0 && |
91327a77 | 1227 | dir->is_auth() && |
11fdf7f2 | 1228 | !dir->inode->is_base() && // not root/mdsdir (for now at least) |
91327a77 | 1229 | !dir->inode->is_stray()) { // not straydir |
7c673cae FG |
1230 | |
1231 | // split | |
f67539c2 | 1232 | if (dir->should_split() || hot) { |
7c673cae FG |
1233 | if (split_pending.count(dir->dirfrag()) == 0) { |
1234 | queue_split(dir, false); | |
1235 | } else { | |
1236 | if (dir->should_split_fast()) { | |
1237 | queue_split(dir, true); | |
1238 | } else { | |
9f95a23c | 1239 | dout(10) << ": fragment already enqueued to split: " |
7c673cae FG |
1240 | << *dir << dendl; |
1241 | } | |
1242 | } | |
1243 | } | |
1244 | ||
1245 | // merge? | |
1e59de90 | 1246 | if (dir->should_merge() && merge_pending.count(dir->dirfrag()) == 0) { |
7c673cae FG |
1247 | queue_merge(dir); |
1248 | } | |
1249 | } | |
1250 | } | |
1251 | ||
1e59de90 | 1252 | void MDBalancer::hit_dir(CDir *dir, int type, double amount) |
7c673cae | 1253 | { |
f67539c2 TL |
1254 | if (dir->inode->is_stray()) |
1255 | return; | |
7c673cae | 1256 | // hit me |
11fdf7f2 | 1257 | double v = dir->pop_me.get(type).hit(amount); |
7c673cae | 1258 | |
11fdf7f2 TL |
1259 | const bool hot = (v > g_conf()->mds_bal_split_rd && type == META_POP_IRD) || |
1260 | (v > g_conf()->mds_bal_split_wr && type == META_POP_IWR); | |
7c673cae | 1261 | |
9f95a23c | 1262 | dout(20) << type << " pop is " << v << ", frag " << dir->get_frag() |
11fdf7f2 | 1263 | << " size " << dir->get_frag_size() << " " << dir->pop_me << dendl; |
7c673cae FG |
1264 | |
1265 | maybe_fragment(dir, hot); | |
1266 | ||
1267 | // replicate? | |
1e59de90 | 1268 | const bool readop = (type == META_POP_IRD || type == META_POP_READDIR); |
7c673cae | 1269 | double rd_adj = 0.0; |
1e59de90 | 1270 | if (readop && dir->last_popularity_sample < last_sample) { |
11fdf7f2 | 1271 | double dir_pop = dir->pop_auth_subtree.get(type).get(); // hmm?? |
1e59de90 | 1272 | dir_pop += v * 10; |
7c673cae | 1273 | dir->last_popularity_sample = last_sample; |
7c673cae | 1274 | |
1e59de90 TL |
1275 | dout(20) << type << " pop " << dir_pop << " spread in " << *dir << dendl; |
1276 | if (dir->is_auth() && !dir->is_ambiguous_auth() && dir->can_rep()) { | |
1277 | if (dir_pop >= g_conf()->mds_bal_replicate_threshold) { | |
7c673cae | 1278 | // replicate |
11fdf7f2 | 1279 | double rdp = dir->pop_me.get(META_POP_IRD).get(); |
7c673cae FG |
1280 | rd_adj = rdp / mds->get_mds_map()->get_num_in_mds() - rdp; |
1281 | rd_adj /= 2.0; // temper somewhat | |
1282 | ||
b32b8144 | 1283 | dout(5) << "replicating dir " << *dir << " pop " << dir_pop << " .. rdp " << rdp << " adj " << rd_adj << dendl; |
7c673cae FG |
1284 | |
1285 | dir->dir_rep = CDir::REP_ALL; | |
1286 | mds->mdcache->send_dir_updates(dir, true); | |
1287 | ||
1288 | // fixme this should adjust the whole pop hierarchy | |
1289 | dir->pop_me.get(META_POP_IRD).adjust(rd_adj); | |
1290 | dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); | |
1291 | } | |
1292 | ||
1293 | if (dir->ino() != 1 && | |
1294 | dir->is_rep() && | |
11fdf7f2 | 1295 | dir_pop < g_conf()->mds_bal_unreplicate_threshold) { |
7c673cae | 1296 | // unreplicate |
b32b8144 | 1297 | dout(5) << "unreplicating dir " << *dir << " pop " << dir_pop << dendl; |
7c673cae FG |
1298 | |
1299 | dir->dir_rep = CDir::REP_NONE; | |
1300 | mds->mdcache->send_dir_updates(dir); | |
1301 | } | |
1302 | } | |
1303 | } | |
1304 | ||
1305 | // adjust ancestors | |
1306 | bool hit_subtree = dir->is_auth(); // current auth subtree (if any) | |
1307 | bool hit_subtree_nested = dir->is_auth(); // all nested auth subtrees | |
1308 | ||
1309 | while (true) { | |
28e407b8 | 1310 | CDir *pdir = dir->inode->get_parent_dir(); |
11fdf7f2 | 1311 | dir->pop_nested.get(type).hit(amount); |
7c673cae | 1312 | if (rd_adj != 0.0) |
11fdf7f2 | 1313 | dir->pop_nested.get(META_POP_IRD).adjust(rd_adj); |
7c673cae FG |
1314 | |
1315 | if (hit_subtree) { | |
11fdf7f2 | 1316 | dir->pop_auth_subtree.get(type).hit(amount); |
28e407b8 | 1317 | |
7c673cae | 1318 | if (rd_adj != 0.0) |
11fdf7f2 | 1319 | dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); |
28e407b8 AA |
1320 | |
1321 | if (dir->is_subtree_root()) | |
1322 | hit_subtree = false; // end of auth domain, stop hitting auth counters. | |
1323 | else if (pdir) | |
1324 | pdir->pop_lru_subdirs.push_front(&dir->get_inode()->item_pop_lru); | |
7c673cae FG |
1325 | } |
1326 | ||
1327 | if (hit_subtree_nested) { | |
11fdf7f2 | 1328 | dir->pop_auth_subtree_nested.get(type).hit(amount); |
7c673cae | 1329 | if (rd_adj != 0.0) |
11fdf7f2 | 1330 | dir->pop_auth_subtree_nested.get(META_POP_IRD).adjust(rd_adj); |
7c673cae | 1331 | } |
28e407b8 AA |
1332 | if (!pdir) break; |
1333 | dir = pdir; | |
7c673cae FG |
1334 | } |
1335 | } | |
1336 | ||
1337 | ||
1338 | /* | |
1339 | * subtract off an exported chunk. | |
1340 | * this excludes *dir itself (encode_export_dir should have take care of that) | |
1341 | * we _just_ do the parents' nested counters. | |
1342 | * | |
1343 | * NOTE: call me _after_ forcing *dir into a subtree root, | |
1344 | * but _before_ doing the encode_export_dirs. | |
1345 | */ | |
11fdf7f2 | 1346 | void MDBalancer::subtract_export(CDir *dir) |
7c673cae FG |
1347 | { |
1348 | dirfrag_load_vec_t subload = dir->pop_auth_subtree; | |
1349 | ||
1350 | while (true) { | |
1351 | dir = dir->inode->get_parent_dir(); | |
1352 | if (!dir) break; | |
1353 | ||
11fdf7f2 TL |
1354 | dir->pop_nested.sub(subload); |
1355 | dir->pop_auth_subtree_nested.sub(subload); | |
7c673cae FG |
1356 | } |
1357 | } | |
1358 | ||
1359 | ||
11fdf7f2 | 1360 | void MDBalancer::add_import(CDir *dir) |
7c673cae FG |
1361 | { |
1362 | dirfrag_load_vec_t subload = dir->pop_auth_subtree; | |
1363 | ||
1364 | while (true) { | |
1365 | dir = dir->inode->get_parent_dir(); | |
1366 | if (!dir) break; | |
1367 | ||
11fdf7f2 TL |
1368 | dir->pop_nested.add(subload); |
1369 | dir->pop_auth_subtree_nested.add(subload); | |
7c673cae FG |
1370 | } |
1371 | } | |
1372 | ||
11fdf7f2 | 1373 | void MDBalancer::adjust_pop_for_rename(CDir *pdir, CDir *dir, bool inc) |
28e407b8 | 1374 | { |
28e407b8 AA |
1375 | bool adjust_subtree_nest = dir->is_auth(); |
1376 | bool adjust_subtree = adjust_subtree_nest && !dir->is_subtree_root(); | |
1377 | CDir *cur = dir; | |
1378 | while (true) { | |
1379 | if (inc) { | |
11fdf7f2 | 1380 | pdir->pop_nested.add(dir->pop_nested); |
28e407b8 | 1381 | if (adjust_subtree) { |
11fdf7f2 | 1382 | pdir->pop_auth_subtree.add(dir->pop_auth_subtree); |
28e407b8 AA |
1383 | pdir->pop_lru_subdirs.push_front(&cur->get_inode()->item_pop_lru); |
1384 | } | |
1385 | ||
1386 | if (adjust_subtree_nest) | |
11fdf7f2 | 1387 | pdir->pop_auth_subtree_nested.add(dir->pop_auth_subtree_nested); |
28e407b8 | 1388 | } else { |
11fdf7f2 | 1389 | pdir->pop_nested.sub(dir->pop_nested); |
28e407b8 | 1390 | if (adjust_subtree) |
11fdf7f2 | 1391 | pdir->pop_auth_subtree.sub(dir->pop_auth_subtree); |
28e407b8 AA |
1392 | |
1393 | if (adjust_subtree_nest) | |
11fdf7f2 | 1394 | pdir->pop_auth_subtree_nested.sub(dir->pop_auth_subtree_nested); |
28e407b8 AA |
1395 | } |
1396 | ||
1397 | if (pdir->is_subtree_root()) | |
1398 | adjust_subtree = false; | |
1399 | cur = pdir; | |
1400 | pdir = pdir->inode->get_parent_dir(); | |
1401 | if (!pdir) break; | |
1402 | } | |
1403 | } | |
1404 | ||
224ce89b WB |
1405 | void MDBalancer::handle_mds_failure(mds_rank_t who) |
1406 | { | |
1407 | if (0 == who) { | |
28e407b8 | 1408 | mds_last_epoch_under_map.clear(); |
224ce89b WB |
1409 | } |
1410 | } | |
28e407b8 | 1411 | |
1e59de90 | 1412 | int MDBalancer::dump_loads(Formatter *f, int64_t depth) const |
28e407b8 | 1413 | { |
1e59de90 TL |
1414 | std::deque<pair<CDir*, int>> dfs; |
1415 | std::deque<CDir*> dfs_root; | |
28e407b8 | 1416 | if (mds->mdcache->get_root()) { |
1e59de90 TL |
1417 | mds->mdcache->get_root()->get_dirfrags(dfs_root); |
1418 | while (!dfs_root.empty()) { | |
1419 | CDir *dir = dfs_root.front(); | |
1420 | dfs_root.pop_front(); | |
1421 | dfs.push_back(make_pair(dir, 0)); | |
1422 | } | |
28e407b8 | 1423 | } else { |
9f95a23c | 1424 | dout(10) << "no root" << dendl; |
28e407b8 AA |
1425 | } |
1426 | ||
1427 | f->open_object_section("loads"); | |
1428 | ||
1429 | f->open_array_section("dirfrags"); | |
1430 | while (!dfs.empty()) { | |
1e59de90 | 1431 | auto [dir, cur_depth] = dfs.front(); |
28e407b8 AA |
1432 | dfs.pop_front(); |
1433 | ||
11fdf7f2 TL |
1434 | f->open_object_section("dir"); |
1435 | dir->dump_load(f); | |
1436 | f->close_section(); | |
28e407b8 | 1437 | |
1e59de90 TL |
1438 | //limit output dirfrags depth |
1439 | if (depth >= 0 && (cur_depth + 1) > depth) { | |
1440 | continue; | |
1441 | } | |
1442 | ||
28e407b8 AA |
1443 | for (auto it = dir->begin(); it != dir->end(); ++it) { |
1444 | CInode *in = it->second->get_linkage()->get_inode(); | |
1445 | if (!in || !in->is_dir()) | |
1446 | continue; | |
1447 | ||
9f95a23c TL |
1448 | auto&& ls = in->get_dirfrags(); |
1449 | for (const auto& subdir : ls) { | |
1e59de90 | 1450 | |
28e407b8 AA |
1451 | if (subdir->pop_nested.meta_load() < .001) |
1452 | continue; | |
1e59de90 | 1453 | dfs.push_back(make_pair(subdir, cur_depth+1)); |
28e407b8 AA |
1454 | } |
1455 | } | |
1456 | } | |
1457 | f->close_section(); // dirfrags array | |
1458 | ||
1459 | f->open_object_section("mds_load"); | |
1460 | { | |
1461 | ||
f67539c2 | 1462 | auto dump_mds_load = [f](const mds_load_t& load) { |
28e407b8 AA |
1463 | f->dump_float("request_rate", load.req_rate); |
1464 | f->dump_float("cache_hit_rate", load.cache_hit_rate); | |
1465 | f->dump_float("queue_length", load.queue_len); | |
1466 | f->dump_float("cpu_load", load.cpu_load_avg); | |
1467 | f->dump_float("mds_load", load.mds_load()); | |
1468 | ||
28e407b8 | 1469 | f->open_object_section("auth_dirfrags"); |
11fdf7f2 | 1470 | load.auth.dump(f); |
28e407b8 AA |
1471 | f->close_section(); |
1472 | f->open_object_section("all_dirfrags"); | |
11fdf7f2 | 1473 | load.all.dump(f); |
28e407b8 AA |
1474 | f->close_section(); |
1475 | }; | |
1476 | ||
f67539c2 TL |
1477 | for (const auto& [rank, load] : mds_load) { |
1478 | CachedStackStringStream css; | |
1479 | *css << "mds." << rank; | |
1480 | f->open_object_section(css->strv()); | |
1481 | dump_mds_load(load); | |
28e407b8 AA |
1482 | f->close_section(); |
1483 | } | |
1484 | } | |
1485 | f->close_section(); // mds_load | |
1486 | ||
1487 | f->open_object_section("mds_meta_load"); | |
f67539c2 TL |
1488 | for (auto& [rank, mload] : mds_meta_load) { |
1489 | CachedStackStringStream css; | |
1490 | *css << "mds." << rank; | |
1491 | f->dump_float(css->strv(), mload); | |
28e407b8 AA |
1492 | } |
1493 | f->close_section(); // mds_meta_load | |
1494 | ||
1495 | f->open_object_section("mds_import_map"); | |
f67539c2 TL |
1496 | for (auto& [rank, imports] : mds_import_map) { |
1497 | { | |
1498 | CachedStackStringStream css; | |
1499 | *css << "mds." << rank; | |
1500 | f->open_array_section(css->strv()); | |
1501 | } | |
1502 | for (auto& [rank_from, mload] : imports) { | |
28e407b8 | 1503 | f->open_object_section("from"); |
f67539c2 TL |
1504 | CachedStackStringStream css; |
1505 | *css << "mds." << rank_from; | |
1506 | f->dump_float(css->strv(), mload); | |
28e407b8 AA |
1507 | f->close_section(); |
1508 | } | |
1509 | f->close_section(); // mds.? array | |
1510 | } | |
1511 | f->close_section(); // mds_import_map | |
1512 | ||
1513 | f->close_section(); // loads | |
1514 | return 0; | |
1515 | } |