]>
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> | |
32 | #include <iostream> | |
33 | #include <vector> | |
34 | #include <map> | |
35 | using std::map; | |
36 | using std::vector; | |
11fdf7f2 | 37 | using std::chrono::duration_cast; |
7c673cae FG |
38 | |
39 | #include "common/config.h" | |
40 | #include "common/errno.h" | |
41 | ||
42 | #define dout_context g_ceph_context | |
7c673cae | 43 | #undef dout_prefix |
9f95a23c | 44 | #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".bal " << __func__ << " " |
7c673cae FG |
45 | #undef dout |
46 | #define dout(lvl) \ | |
47 | do {\ | |
48 | auto subsys = ceph_subsys_mds;\ | |
49 | if ((dout_context)->_conf->subsys.should_gather(ceph_subsys_mds_balancer, lvl)) {\ | |
50 | subsys = ceph_subsys_mds_balancer;\ | |
51 | }\ | |
11fdf7f2 | 52 | dout_impl(dout_context, ceph::dout::need_dynamic(subsys), lvl) dout_prefix |
7c673cae FG |
53 | #undef dendl |
54 | #define dendl dendl_impl; } while (0) | |
55 | ||
56 | ||
57 | #define MIN_LOAD 50 // ?? | |
58 | #define MIN_REEXPORT 5 // will automatically reexport | |
59 | #define MIN_OFFLOAD 10 // point at which i stop trying, close enough | |
60 | ||
61 | ||
9f95a23c | 62 | int MDBalancer::proc_message(const cref_t<Message> &m) |
7c673cae FG |
63 | { |
64 | switch (m->get_type()) { | |
65 | ||
66 | case MSG_MDS_HEARTBEAT: | |
9f95a23c | 67 | handle_heartbeat(ref_cast<MHeartbeat>(m)); |
7c673cae FG |
68 | break; |
69 | ||
70 | default: | |
b32b8144 | 71 | derr << " balancer unknown message " << m->get_type() << dendl_impl; |
11fdf7f2 | 72 | ceph_abort_msg("balancer unknown message"); |
7c673cae FG |
73 | } |
74 | ||
75 | return 0; | |
76 | } | |
77 | ||
91327a77 AA |
78 | MDBalancer::MDBalancer(MDSRank *m, Messenger *msgr, MonClient *monc) : |
79 | mds(m), messenger(msgr), mon_client(monc) | |
80 | { | |
11fdf7f2 TL |
81 | bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); |
82 | bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); | |
91327a77 AA |
83 | } |
84 | ||
92f5a8d4 | 85 | void MDBalancer::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mds_map) |
91327a77 | 86 | { |
11fdf7f2 TL |
87 | if (changed.count("mds_bal_fragment_dirs")) |
88 | bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); | |
91327a77 | 89 | if (changed.count("mds_bal_fragment_interval")) |
11fdf7f2 | 90 | bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); |
91327a77 AA |
91 | } |
92 | ||
7c673cae FG |
93 | void MDBalancer::handle_export_pins(void) |
94 | { | |
95 | auto &q = mds->mdcache->export_pin_queue; | |
96 | auto it = q.begin(); | |
97 | dout(20) << "export_pin_queue size=" << q.size() << dendl; | |
98 | while (it != q.end()) { | |
31f18b77 FG |
99 | auto cur = it++; |
100 | CInode *in = *cur; | |
11fdf7f2 | 101 | ceph_assert(in->is_dir()); |
f6b5b4d7 TL |
102 | |
103 | in->check_pin_policy(); | |
31f18b77 | 104 | mds_rank_t export_pin = in->get_export_pin(false); |
eafe8130 | 105 | if (export_pin >= mds->mdsmap->get_max_mds()) { |
f6b5b4d7 | 106 | dout(20) << " delay export_pin=" << export_pin << " on " << *in << dendl; |
eafe8130 TL |
107 | in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); |
108 | q.erase(cur); | |
109 | ||
110 | in->state_set(CInode::STATE_DELAYEDEXPORTPIN); | |
111 | mds->mdcache->export_pin_delayed_queue.insert(in); | |
112 | continue; | |
f6b5b4d7 TL |
113 | } else { |
114 | dout(20) << " executing export_pin=" << export_pin << " on " << *in << dendl; | |
eafe8130 | 115 | } |
31f18b77 FG |
116 | |
117 | bool remove = true; | |
f6b5b4d7 | 118 | for (auto&& dir : in->get_dirfrags()) { |
31f18b77 FG |
119 | if (!dir->is_auth()) |
120 | continue; | |
121 | ||
122 | if (export_pin == MDS_RANK_NONE) { | |
123 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) { | |
124 | if (dir->is_frozen() || dir->is_freezing()) { | |
125 | // try again later | |
126 | remove = false; | |
127 | continue; | |
128 | } | |
129 | dout(10) << " clear auxsubtree on " << *dir << dendl; | |
130 | dir->state_clear(CDir::STATE_AUXSUBTREE); | |
131 | mds->mdcache->try_subtree_merge(dir); | |
132 | } | |
133 | } else if (export_pin == mds->get_nodeid()) { | |
f6b5b4d7 TL |
134 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) { |
135 | ceph_assert(dir->is_subtree_root()); | |
136 | } else if (dir->state_test(CDir::STATE_CREATING) || | |
137 | dir->is_frozen() || dir->is_freezing()) { | |
31f18b77 FG |
138 | // try again later |
139 | remove = false; | |
140 | continue; | |
f6b5b4d7 | 141 | } else if (!dir->is_subtree_root()) { |
31f18b77 FG |
142 | dir->state_set(CDir::STATE_AUXSUBTREE); |
143 | mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid()); | |
144 | dout(10) << " create aux subtree on " << *dir << dendl; | |
f6b5b4d7 | 145 | } else { |
31f18b77 FG |
146 | dout(10) << " set auxsubtree bit on " << *dir << dendl; |
147 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
148 | } | |
149 | } else { | |
f6b5b4d7 TL |
150 | /* Only export a directory if it's non-empty. An empty directory will |
151 | * be sent back by the importer. | |
152 | */ | |
153 | if (dir->get_num_head_items() > 0) { | |
154 | mds->mdcache->migrator->export_dir(dir, export_pin); | |
155 | } | |
31f18b77 | 156 | remove = false; |
7c673cae FG |
157 | } |
158 | } | |
31f18b77 FG |
159 | |
160 | if (remove) { | |
161 | in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); | |
162 | q.erase(cur); | |
7c673cae FG |
163 | } |
164 | } | |
165 | ||
81eedcae TL |
166 | std::vector<CDir *> authsubs = mds->mdcache->get_auth_subtrees(); |
167 | bool print_auth_subtrees = true; | |
168 | ||
169 | if (authsubs.size() > AUTH_TREES_THRESHOLD && | |
170 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
171 | dout(15) << "number of auth trees = " << authsubs.size() << "; not " | |
172 | "printing auth trees" << dendl; | |
173 | print_auth_subtrees = false; | |
174 | } | |
175 | ||
176 | for (auto &cd : authsubs) { | |
7c673cae | 177 | mds_rank_t export_pin = cd->inode->get_export_pin(); |
81eedcae TL |
178 | |
179 | if (print_auth_subtrees) { | |
180 | dout(25) << "auth tree " << *cd << " export_pin=" << export_pin << | |
181 | dendl; | |
182 | } | |
183 | ||
f6b5b4d7 TL |
184 | if (export_pin >= 0 && export_pin < mds->mdsmap->get_max_mds()) { |
185 | if (export_pin == mds->get_nodeid()) { | |
186 | cd->get_inode()->check_pin_policy(); | |
187 | } else { | |
188 | mds->mdcache->migrator->export_dir(cd, export_pin); | |
189 | } | |
7c673cae FG |
190 | } |
191 | } | |
192 | } | |
193 | ||
194 | void MDBalancer::tick() | |
195 | { | |
11fdf7f2 TL |
196 | static int num_bal_times = g_conf()->mds_bal_max; |
197 | auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); | |
198 | auto bal_max_until = g_conf().get_val<int64_t>("mds_bal_max_until"); | |
199 | time now = clock::now(); | |
7c673cae | 200 | |
11fdf7f2 | 201 | if (g_conf()->mds_bal_export_pin) { |
7c673cae FG |
202 | handle_export_pins(); |
203 | } | |
204 | ||
205 | // sample? | |
11fdf7f2 TL |
206 | if (chrono::duration<double>(now-last_sample).count() > |
207 | g_conf()->mds_bal_sample_interval) { | |
7c673cae FG |
208 | dout(15) << "tick last_sample now " << now << dendl; |
209 | last_sample = now; | |
210 | } | |
211 | ||
11fdf7f2 TL |
212 | // We can use duration_cast below, although the result is an int, |
213 | // because the values from g_conf are also integers. | |
7c673cae | 214 | // balance? |
11fdf7f2 TL |
215 | if (mds->get_nodeid() == 0 |
216 | && mds->is_active() | |
217 | && bal_interval > 0 | |
218 | && duration_cast<chrono::seconds>(now - last_heartbeat).count() >= bal_interval | |
219 | && (num_bal_times || (bal_max_until >= 0 && mds->get_uptime().count() > bal_max_until))) { | |
7c673cae FG |
220 | last_heartbeat = now; |
221 | send_heartbeat(); | |
222 | num_bal_times--; | |
223 | } | |
81eedcae TL |
224 | |
225 | mds->mdcache->show_subtrees(10, true); | |
7c673cae FG |
226 | } |
227 | ||
228 | ||
229 | ||
230 | ||
231 | class C_Bal_SendHeartbeat : public MDSInternalContext { | |
232 | public: | |
233 | explicit C_Bal_SendHeartbeat(MDSRank *mds_) : MDSInternalContext(mds_) { } | |
234 | void finish(int f) override { | |
235 | mds->balancer->send_heartbeat(); | |
236 | } | |
237 | }; | |
238 | ||
239 | ||
11fdf7f2 | 240 | double mds_load_t::mds_load() const |
7c673cae | 241 | { |
11fdf7f2 | 242 | switch(g_conf()->mds_bal_mode) { |
7c673cae FG |
243 | case 0: |
244 | return | |
245 | .8 * auth.meta_load() + | |
246 | .2 * all.meta_load() + | |
247 | req_rate + | |
248 | 10.0 * queue_len; | |
249 | ||
250 | case 1: | |
251 | return req_rate + 10.0*queue_len; | |
252 | ||
253 | case 2: | |
254 | return cpu_load_avg; | |
255 | ||
256 | } | |
257 | ceph_abort(); | |
258 | return 0; | |
259 | } | |
260 | ||
11fdf7f2 | 261 | mds_load_t MDBalancer::get_load() |
7c673cae | 262 | { |
11fdf7f2 TL |
263 | auto now = clock::now(); |
264 | ||
265 | mds_load_t load{DecayRate()}; /* zero DecayRate! */ | |
7c673cae FG |
266 | |
267 | if (mds->mdcache->get_root()) { | |
9f95a23c | 268 | auto&& ls = mds->mdcache->get_root()->get_dirfrags(); |
11fdf7f2 TL |
269 | for (auto &d : ls) { |
270 | load.auth.add(d->pop_auth_subtree_nested); | |
271 | load.all.add(d->pop_nested); | |
7c673cae FG |
272 | } |
273 | } else { | |
9f95a23c | 274 | dout(20) << "no root, no load" << dendl; |
7c673cae FG |
275 | } |
276 | ||
28e407b8 | 277 | uint64_t num_requests = mds->get_num_requests(); |
91327a77 AA |
278 | |
279 | uint64_t cpu_time = 1; | |
280 | { | |
281 | string stat_path = PROCPREFIX "/proc/self/stat"; | |
282 | ifstream stat_file(stat_path); | |
283 | if (stat_file.is_open()) { | |
284 | vector<string> stat_vec(std::istream_iterator<string>{stat_file}, | |
285 | std::istream_iterator<string>()); | |
286 | if (stat_vec.size() >= 15) { | |
287 | // utime + stime | |
288 | cpu_time = strtoll(stat_vec[13].c_str(), nullptr, 10) + | |
289 | strtoll(stat_vec[14].c_str(), nullptr, 10); | |
290 | } else { | |
291 | derr << "input file '" << stat_path << "' not resolvable" << dendl_impl; | |
292 | } | |
293 | } else { | |
294 | derr << "input file '" << stat_path << "' not found" << dendl_impl; | |
295 | } | |
296 | } | |
297 | ||
298 | load.queue_len = messenger->get_dispatch_queue_len(); | |
299 | ||
300 | bool update_last = true; | |
11fdf7f2 | 301 | if (last_get_load != clock::zero() && |
91327a77 | 302 | now > last_get_load) { |
11fdf7f2 TL |
303 | double el = std::chrono::duration<double>(now-last_get_load).count(); |
304 | if (el >= 1.0) { | |
91327a77 | 305 | if (num_requests > last_num_requests) |
11fdf7f2 | 306 | load.req_rate = (num_requests - last_num_requests) / el; |
91327a77 | 307 | if (cpu_time > last_cpu_time) |
11fdf7f2 | 308 | load.cpu_load_avg = (cpu_time - last_cpu_time) / el; |
91327a77 AA |
309 | } else { |
310 | auto p = mds_load.find(mds->get_nodeid()); | |
311 | if (p != mds_load.end()) { | |
312 | load.req_rate = p->second.req_rate; | |
313 | load.cpu_load_avg = p->second.cpu_load_avg; | |
314 | } | |
315 | if (num_requests >= last_num_requests && cpu_time >= last_cpu_time) | |
316 | update_last = false; | |
28e407b8 AA |
317 | } |
318 | } | |
28e407b8 | 319 | |
91327a77 AA |
320 | if (update_last) { |
321 | last_num_requests = num_requests; | |
322 | last_cpu_time = cpu_time; | |
323 | last_get_load = now; | |
324 | } | |
7c673cae | 325 | |
9f95a23c | 326 | dout(15) << load << dendl; |
7c673cae FG |
327 | return load; |
328 | } | |
329 | ||
330 | /* | |
331 | * Read synchronously from RADOS using a timeout. We cannot do daemon-local | |
332 | * fallbacks (i.e. kick off async read when we are processing the map and | |
333 | * check status when we get here) with the way the mds is structured. | |
334 | */ | |
335 | int MDBalancer::localize_balancer() | |
336 | { | |
337 | /* reset everything */ | |
338 | bool ack = false; | |
339 | int r = 0; | |
340 | bufferlist lua_src; | |
9f95a23c TL |
341 | ceph::mutex lock = ceph::make_mutex("lock"); |
342 | ceph::condition_variable cond; | |
7c673cae FG |
343 | |
344 | /* we assume that balancer is in the metadata pool */ | |
345 | object_t oid = object_t(mds->mdsmap->get_balancer()); | |
346 | object_locator_t oloc(mds->mdsmap->get_metadata_pool()); | |
347 | ceph_tid_t tid = mds->objecter->read(oid, oloc, 0, 0, CEPH_NOSNAP, &lua_src, 0, | |
9f95a23c | 348 | new C_SafeCond(lock, cond, &ack, &r)); |
7c673cae FG |
349 | dout(15) << "launched non-blocking read tid=" << tid |
350 | << " oid=" << oid << " oloc=" << oloc << dendl; | |
351 | ||
352 | /* timeout: if we waste half our time waiting for RADOS, then abort! */ | |
9f95a23c TL |
353 | std::cv_status ret_t = [&] { |
354 | auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); | |
355 | std::unique_lock locker{lock}; | |
356 | return cond.wait_for(locker, std::chrono::seconds(bal_interval / 2)); | |
357 | }(); | |
7c673cae FG |
358 | /* success: store the balancer in memory and set the version. */ |
359 | if (!r) { | |
9f95a23c | 360 | if (ret_t == std::cv_status::timeout) { |
7c673cae FG |
361 | mds->objecter->op_cancel(tid, -ECANCELED); |
362 | return -ETIMEDOUT; | |
363 | } | |
364 | bal_code.assign(lua_src.to_str()); | |
365 | bal_version.assign(oid.name); | |
9f95a23c | 366 | dout(10) "bal_code=" << bal_code << dendl; |
7c673cae FG |
367 | } |
368 | return r; | |
369 | } | |
370 | ||
371 | void MDBalancer::send_heartbeat() | |
372 | { | |
7c673cae | 373 | if (mds->is_cluster_degraded()) { |
9f95a23c | 374 | dout(10) << "degraded" << dendl; |
7c673cae FG |
375 | return; |
376 | } | |
377 | ||
378 | if (!mds->mdcache->is_open()) { | |
9f95a23c | 379 | dout(10) << "not open" << dendl; |
7c673cae FG |
380 | mds->mdcache->wait_for_open(new C_Bal_SendHeartbeat(mds)); |
381 | return; | |
382 | } | |
383 | ||
94b18763 | 384 | if (mds->get_nodeid() == 0) { |
7c673cae | 385 | beat_epoch++; |
94b18763 FG |
386 | mds_load.clear(); |
387 | } | |
7c673cae FG |
388 | |
389 | // my load | |
11fdf7f2 | 390 | mds_load_t load = get_load(); |
28e407b8 AA |
391 | mds->logger->set(l_mds_load_cent, 100 * load.mds_load()); |
392 | mds->logger->set(l_mds_dispatch_queue_len, load.queue_len); | |
393 | ||
11fdf7f2 TL |
394 | auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(mds->get_nodeid()), std::forward_as_tuple(load)); |
395 | if (!em.second) { | |
396 | em.first->second = load; | |
397 | } | |
7c673cae FG |
398 | |
399 | // import_map -- how much do i import from whom | |
400 | map<mds_rank_t, float> import_map; | |
11fdf7f2 | 401 | for (auto& im : mds->mdcache->get_auth_subtrees()) { |
7c673cae FG |
402 | mds_rank_t from = im->inode->authority().first; |
403 | if (from == mds->get_nodeid()) continue; | |
404 | if (im->get_inode()->is_stray()) continue; | |
11fdf7f2 | 405 | import_map[from] += im->pop_auth_subtree.meta_load(); |
7c673cae FG |
406 | } |
407 | mds_import_map[ mds->get_nodeid() ] = import_map; | |
408 | ||
409 | ||
9f95a23c TL |
410 | dout(3) << " epoch " << beat_epoch << " load " << load << dendl; |
411 | for (const auto& [rank, load] : import_map) { | |
412 | dout(5) << " import_map from " << rank << " -> " << load << dendl; | |
7c673cae FG |
413 | } |
414 | ||
415 | ||
416 | set<mds_rank_t> up; | |
417 | mds->get_mds_map()->get_up_mds_set(up); | |
11fdf7f2 TL |
418 | for (const auto& r : up) { |
419 | if (r == mds->get_nodeid()) | |
7c673cae | 420 | continue; |
9f95a23c | 421 | auto hb = make_message<MHeartbeat>(load, beat_epoch); |
7c673cae | 422 | hb->get_import_map() = import_map; |
11fdf7f2 | 423 | mds->send_message_mds(hb, r); |
7c673cae FG |
424 | } |
425 | } | |
426 | ||
9f95a23c | 427 | void MDBalancer::handle_heartbeat(const cref_t<MHeartbeat> &m) |
7c673cae | 428 | { |
7c673cae FG |
429 | mds_rank_t who = mds_rank_t(m->get_source().num()); |
430 | dout(25) << "=== got heartbeat " << m->get_beat() << " from " << m->get_source().num() << " " << m->get_load() << dendl; | |
431 | ||
432 | if (!mds->is_active()) | |
11fdf7f2 | 433 | return; |
7c673cae FG |
434 | |
435 | if (!mds->mdcache->is_open()) { | |
436 | dout(10) << "opening root on handle_heartbeat" << dendl; | |
437 | mds->mdcache->wait_for_open(new C_MDS_RetryMessage(mds, m)); | |
438 | return; | |
439 | } | |
440 | ||
441 | if (mds->is_cluster_degraded()) { | |
442 | dout(10) << " degraded, ignoring" << dendl; | |
11fdf7f2 | 443 | return; |
7c673cae FG |
444 | } |
445 | ||
94b18763 FG |
446 | if (mds->get_nodeid() != 0 && m->get_beat() > beat_epoch) { |
447 | dout(10) << "receive next epoch " << m->get_beat() << " from mds." << who << " before mds0" << dendl; | |
448 | ||
449 | beat_epoch = m->get_beat(); | |
450 | // clear the mds load info whose epoch is less than beat_epoch | |
451 | mds_load.clear(); | |
452 | } | |
453 | ||
7c673cae | 454 | if (who == 0) { |
94b18763 FG |
455 | dout(20) << " from mds0, new epoch " << m->get_beat() << dendl; |
456 | if (beat_epoch != m->get_beat()) { | |
28e407b8 | 457 | beat_epoch = m->get_beat(); |
94b18763 FG |
458 | mds_load.clear(); |
459 | } | |
28e407b8 | 460 | |
7c673cae FG |
461 | send_heartbeat(); |
462 | ||
463 | mds->mdcache->show_subtrees(); | |
28e407b8 AA |
464 | } else if (mds->get_nodeid() == 0) { |
465 | if (beat_epoch != m->get_beat()) { | |
466 | dout(10) << " old heartbeat epoch, ignoring" << dendl; | |
11fdf7f2 | 467 | return; |
7c673cae FG |
468 | } |
469 | } | |
28e407b8 | 470 | |
11fdf7f2 TL |
471 | { |
472 | auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(who), std::forward_as_tuple(m->get_load())); | |
473 | if (!em.second) { | |
474 | em.first->second = m->get_load(); | |
475 | } | |
476 | } | |
28e407b8 | 477 | mds_import_map[who] = m->get_import_map(); |
7c673cae | 478 | |
7c673cae FG |
479 | { |
480 | unsigned cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
481 | if (mds_load.size() == cluster_size) { | |
482 | // let's go! | |
483 | //export_empties(); // no! | |
484 | ||
485 | /* avoid spamming ceph -w if user does not turn mantle on */ | |
486 | if (mds->mdsmap->get_balancer() != "") { | |
487 | int r = mantle_prep_rebalance(); | |
11fdf7f2 | 488 | if (!r) return; |
7c673cae FG |
489 | mds->clog->warn() << "using old balancer; mantle failed for " |
490 | << "balancer=" << mds->mdsmap->get_balancer() | |
491 | << " : " << cpp_strerror(r); | |
492 | } | |
493 | prep_rebalance(m->get_beat()); | |
494 | } | |
495 | } | |
7c673cae FG |
496 | } |
497 | ||
7c673cae FG |
498 | double MDBalancer::try_match(balance_state_t& state, mds_rank_t ex, double& maxex, |
499 | mds_rank_t im, double& maxim) | |
500 | { | |
501 | if (maxex <= 0 || maxim <= 0) return 0.0; | |
502 | ||
11fdf7f2 | 503 | double howmuch = std::min(maxex, maxim); |
7c673cae FG |
504 | |
505 | dout(5) << " - mds." << ex << " exports " << howmuch << " to mds." << im << dendl; | |
506 | ||
507 | if (ex == mds->get_nodeid()) | |
508 | state.targets[im] += howmuch; | |
509 | ||
510 | state.exported[ex] += howmuch; | |
511 | state.imported[im] += howmuch; | |
512 | ||
513 | maxex -= howmuch; | |
514 | maxim -= howmuch; | |
515 | ||
516 | return howmuch; | |
517 | } | |
518 | ||
519 | void MDBalancer::queue_split(const CDir *dir, bool fast) | |
520 | { | |
521 | dout(10) << __func__ << " enqueuing " << *dir | |
522 | << " (fast=" << fast << ")" << dendl; | |
523 | ||
7c673cae FG |
524 | const dirfrag_t frag = dir->dirfrag(); |
525 | ||
526 | auto callback = [this, frag](int r) { | |
527 | if (split_pending.erase(frag) == 0) { | |
528 | // Someone beat me to it. This can happen in the fast splitting | |
529 | // path, because we spawn two contexts, one with mds->timer and | |
530 | // one with mds->queue_waiter. The loser can safely just drop | |
531 | // out. | |
532 | return; | |
533 | } | |
534 | ||
535 | CDir *split_dir = mds->mdcache->get_dirfrag(frag); | |
536 | if (!split_dir) { | |
537 | dout(10) << "drop split on " << frag << " because not in cache" << dendl; | |
538 | return; | |
539 | } | |
540 | if (!split_dir->is_auth()) { | |
541 | dout(10) << "drop split on " << frag << " because non-auth" << dendl; | |
542 | return; | |
543 | } | |
544 | ||
545 | // Pass on to MDCache: note that the split might still not | |
546 | // happen if the checks in MDCache::can_fragment fail. | |
547 | dout(10) << __func__ << " splitting " << *split_dir << dendl; | |
11fdf7f2 | 548 | mds->mdcache->split_dir(split_dir, g_conf()->mds_bal_split_bits); |
7c673cae FG |
549 | }; |
550 | ||
551 | bool is_new = false; | |
552 | if (split_pending.count(frag) == 0) { | |
553 | split_pending.insert(frag); | |
554 | is_new = true; | |
555 | } | |
556 | ||
557 | if (fast) { | |
558 | // Do the split ASAP: enqueue it in the MDSRank waiters which are | |
559 | // run at the end of dispatching the current request | |
560 | mds->queue_waiter(new MDSInternalContextWrapper(mds, | |
9f95a23c | 561 | new LambdaContext(std::move(callback)))); |
7c673cae FG |
562 | } else if (is_new) { |
563 | // Set a timer to really do the split: we don't do it immediately | |
564 | // so that bursts of ops on a directory have a chance to go through | |
565 | // before we freeze it. | |
91327a77 | 566 | mds->timer.add_event_after(bal_fragment_interval, |
9f95a23c | 567 | new LambdaContext(std::move(callback))); |
7c673cae FG |
568 | } |
569 | } | |
570 | ||
571 | void MDBalancer::queue_merge(CDir *dir) | |
572 | { | |
573 | const auto frag = dir->dirfrag(); | |
574 | auto callback = [this, frag](int r) { | |
11fdf7f2 | 575 | ceph_assert(frag.frag != frag_t()); |
7c673cae FG |
576 | |
577 | // frag must be in this set because only one context is in flight | |
578 | // for a given frag at a time (because merge_pending is checked before | |
579 | // starting one), and this context is the only one that erases it. | |
580 | merge_pending.erase(frag); | |
581 | ||
582 | CDir *dir = mds->mdcache->get_dirfrag(frag); | |
583 | if (!dir) { | |
584 | dout(10) << "drop merge on " << frag << " because not in cache" << dendl; | |
585 | return; | |
586 | } | |
11fdf7f2 | 587 | ceph_assert(dir->dirfrag() == frag); |
7c673cae FG |
588 | |
589 | if(!dir->is_auth()) { | |
590 | dout(10) << "drop merge on " << *dir << " because lost auth" << dendl; | |
591 | return; | |
592 | } | |
593 | ||
594 | dout(10) << "merging " << *dir << dendl; | |
595 | ||
596 | CInode *diri = dir->get_inode(); | |
597 | ||
598 | frag_t fg = dir->get_frag(); | |
599 | while (fg != frag_t()) { | |
600 | frag_t sibfg = fg.get_sibling(); | |
9f95a23c | 601 | auto&& [complete, sibs] = diri->get_dirfrags_under(sibfg); |
7c673cae FG |
602 | if (!complete) { |
603 | dout(10) << " not all sibs under " << sibfg << " in cache (have " << sibs << ")" << dendl; | |
604 | break; | |
605 | } | |
606 | bool all = true; | |
9f95a23c | 607 | for (auto& sib : sibs) { |
7c673cae FG |
608 | if (!sib->is_auth() || !sib->should_merge()) { |
609 | all = false; | |
610 | break; | |
611 | } | |
612 | } | |
613 | if (!all) { | |
614 | dout(10) << " not all sibs under " << sibfg << " " << sibs << " should_merge" << dendl; | |
615 | break; | |
616 | } | |
617 | dout(10) << " all sibs under " << sibfg << " " << sibs << " should merge" << dendl; | |
618 | fg = fg.parent(); | |
619 | } | |
620 | ||
621 | if (fg != dir->get_frag()) | |
622 | mds->mdcache->merge_dir(diri, fg); | |
623 | }; | |
624 | ||
625 | if (merge_pending.count(frag) == 0) { | |
9f95a23c | 626 | dout(20) << " enqueued dir " << *dir << dendl; |
7c673cae | 627 | merge_pending.insert(frag); |
91327a77 | 628 | mds->timer.add_event_after(bal_fragment_interval, |
9f95a23c | 629 | new LambdaContext(std::move(callback))); |
7c673cae | 630 | } else { |
9f95a23c | 631 | dout(20) << " dir already in queue " << *dir << dendl; |
7c673cae FG |
632 | } |
633 | } | |
634 | ||
635 | void MDBalancer::prep_rebalance(int beat) | |
636 | { | |
637 | balance_state_t state; | |
638 | ||
11fdf7f2 | 639 | if (g_conf()->mds_thrash_exports) { |
7c673cae FG |
640 | //we're going to randomly export to all the mds in the cluster |
641 | set<mds_rank_t> up_mds; | |
642 | mds->get_mds_map()->get_up_mds_set(up_mds); | |
643 | for (const auto &rank : up_mds) { | |
644 | state.targets[rank] = 0.0; | |
645 | } | |
646 | } else { | |
647 | int cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
648 | mds_rank_t whoami = mds->get_nodeid(); | |
11fdf7f2 | 649 | rebalance_time = clock::now(); |
7c673cae | 650 | |
9f95a23c | 651 | dout(7) << "cluster loads are" << dendl; |
7c673cae FG |
652 | |
653 | mds->mdcache->migrator->clear_export_queue(); | |
654 | ||
655 | // rescale! turn my mds_load back into meta_load units | |
656 | double load_fac = 1.0; | |
657 | map<mds_rank_t, mds_load_t>::iterator m = mds_load.find(whoami); | |
658 | if ((m != mds_load.end()) && (m->second.mds_load() > 0)) { | |
11fdf7f2 | 659 | double metald = m->second.auth.meta_load(); |
7c673cae FG |
660 | double mdsld = m->second.mds_load(); |
661 | load_fac = metald / mdsld; | |
662 | dout(7) << " load_fac is " << load_fac | |
663 | << " <- " << m->second.auth << " " << metald | |
664 | << " / " << mdsld | |
665 | << dendl; | |
666 | } | |
667 | ||
28e407b8 AA |
668 | mds_meta_load.clear(); |
669 | ||
7c673cae FG |
670 | double total_load = 0.0; |
671 | multimap<double,mds_rank_t> load_map; | |
672 | for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { | |
28e407b8 | 673 | mds_load_t& load = mds_load.at(i); |
7c673cae FG |
674 | |
675 | double l = load.mds_load() * load_fac; | |
676 | mds_meta_load[i] = l; | |
677 | ||
678 | if (whoami == 0) | |
9f95a23c | 679 | dout(7) << " mds." << i |
7c673cae FG |
680 | << " " << load |
681 | << " = " << load.mds_load() | |
682 | << " ~ " << l << dendl; | |
683 | ||
684 | if (whoami == i) my_load = l; | |
685 | total_load += l; | |
686 | ||
687 | load_map.insert(pair<double,mds_rank_t>( l, i )); | |
688 | } | |
689 | ||
690 | // target load | |
691 | target_load = total_load / (double)cluster_size; | |
9f95a23c | 692 | dout(7) << "my load " << my_load |
7c673cae FG |
693 | << " target " << target_load |
694 | << " total " << total_load | |
695 | << dendl; | |
696 | ||
697 | // under or over? | |
9f95a23c TL |
698 | for (const auto& [load, rank] : load_map) { |
699 | if (load < target_load * (1.0 + g_conf()->mds_bal_min_rebalance)) { | |
700 | dout(7) << " mds." << rank << " is underloaded or barely overloaded." << dendl; | |
701 | mds_last_epoch_under_map[rank] = beat_epoch; | |
28e407b8 AA |
702 | } |
703 | } | |
704 | ||
705 | int last_epoch_under = mds_last_epoch_under_map[whoami]; | |
706 | if (last_epoch_under == beat_epoch) { | |
9f95a23c | 707 | dout(7) << " i am underloaded or barely overloaded, doing nothing." << dendl; |
7c673cae FG |
708 | return; |
709 | } | |
7c673cae FG |
710 | // am i over long enough? |
711 | if (last_epoch_under && beat_epoch - last_epoch_under < 2) { | |
9f95a23c | 712 | dout(7) << " i am overloaded, but only for " << (beat_epoch - last_epoch_under) << " epochs" << dendl; |
7c673cae FG |
713 | return; |
714 | } | |
715 | ||
9f95a23c | 716 | dout(7) << " i am sufficiently overloaded" << dendl; |
7c673cae FG |
717 | |
718 | ||
719 | // first separate exporters and importers | |
720 | multimap<double,mds_rank_t> importers; | |
721 | multimap<double,mds_rank_t> exporters; | |
722 | set<mds_rank_t> importer_set; | |
723 | set<mds_rank_t> exporter_set; | |
724 | ||
725 | for (multimap<double,mds_rank_t>::iterator it = load_map.begin(); | |
726 | it != load_map.end(); | |
727 | ++it) { | |
728 | if (it->first < target_load) { | |
729 | dout(15) << " mds." << it->second << " is importer" << dendl; | |
730 | importers.insert(pair<double,mds_rank_t>(it->first,it->second)); | |
731 | importer_set.insert(it->second); | |
732 | } else { | |
28e407b8 AA |
733 | int mds_last_epoch_under = mds_last_epoch_under_map[it->second]; |
734 | if (!(mds_last_epoch_under && beat_epoch - mds_last_epoch_under < 2)) { | |
735 | dout(15) << " mds." << it->second << " is exporter" << dendl; | |
736 | exporters.insert(pair<double,mds_rank_t>(it->first,it->second)); | |
737 | exporter_set.insert(it->second); | |
738 | } | |
7c673cae FG |
739 | } |
740 | } | |
741 | ||
742 | ||
743 | // determine load transfer mapping | |
744 | ||
745 | if (true) { | |
746 | // analyze import_map; do any matches i can | |
747 | ||
748 | dout(15) << " matching exporters to import sources" << dendl; | |
749 | ||
750 | // big -> small exporters | |
751 | for (multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); | |
752 | ex != exporters.rend(); | |
753 | ++ex) { | |
754 | double maxex = get_maxex(state, ex->second); | |
755 | if (maxex <= .001) continue; | |
756 | ||
757 | // check importers. for now, just in arbitrary order (no intelligent matching). | |
758 | for (map<mds_rank_t, float>::iterator im = mds_import_map[ex->second].begin(); | |
759 | im != mds_import_map[ex->second].end(); | |
760 | ++im) { | |
761 | double maxim = get_maxim(state, im->first); | |
762 | if (maxim <= .001) continue; | |
763 | try_match(state, ex->second, maxex, im->first, maxim); | |
764 | if (maxex <= .001) break; | |
765 | } | |
766 | } | |
767 | } | |
768 | ||
769 | // old way | |
770 | if (beat % 2 == 1) { | |
771 | dout(15) << " matching big exporters to big importers" << dendl; | |
772 | // big exporters to big importers | |
773 | multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); | |
774 | multimap<double,mds_rank_t>::iterator im = importers.begin(); | |
775 | while (ex != exporters.rend() && | |
776 | im != importers.end()) { | |
777 | double maxex = get_maxex(state, ex->second); | |
778 | double maxim = get_maxim(state, im->second); | |
779 | if (maxex < .001 || maxim < .001) break; | |
780 | try_match(state, ex->second, maxex, im->second, maxim); | |
781 | if (maxex <= .001) ++ex; | |
782 | if (maxim <= .001) ++im; | |
783 | } | |
784 | } else { // new way | |
785 | dout(15) << " matching small exporters to big importers" << dendl; | |
786 | // small exporters to big importers | |
787 | multimap<double,mds_rank_t>::iterator ex = exporters.begin(); | |
788 | multimap<double,mds_rank_t>::iterator im = importers.begin(); | |
789 | while (ex != exporters.end() && | |
790 | im != importers.end()) { | |
791 | double maxex = get_maxex(state, ex->second); | |
792 | double maxim = get_maxim(state, im->second); | |
793 | if (maxex < .001 || maxim < .001) break; | |
794 | try_match(state, ex->second, maxex, im->second, maxim); | |
795 | if (maxex <= .001) ++ex; | |
796 | if (maxim <= .001) ++im; | |
797 | } | |
798 | } | |
799 | } | |
800 | try_rebalance(state); | |
801 | } | |
802 | ||
7c673cae FG |
803 | int MDBalancer::mantle_prep_rebalance() |
804 | { | |
805 | balance_state_t state; | |
806 | ||
807 | /* refresh balancer if it has changed */ | |
808 | if (bal_version != mds->mdsmap->get_balancer()) { | |
809 | bal_version.assign(""); | |
810 | int r = localize_balancer(); | |
811 | if (r) return r; | |
812 | ||
813 | /* only spam the cluster log from 1 mds on version changes */ | |
814 | if (mds->get_nodeid() == 0) | |
815 | mds->clog->info() << "mantle balancer version changed: " << bal_version; | |
816 | } | |
817 | ||
818 | /* prepare for balancing */ | |
819 | int cluster_size = mds->get_mds_map()->get_num_in_mds(); | |
11fdf7f2 | 820 | rebalance_time = clock::now(); |
7c673cae FG |
821 | mds->mdcache->migrator->clear_export_queue(); |
822 | ||
823 | /* fill in the metrics for each mds by grabbing load struct */ | |
824 | vector < map<string, double> > metrics (cluster_size); | |
28e407b8 AA |
825 | for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { |
826 | mds_load_t& load = mds_load.at(i); | |
7c673cae FG |
827 | |
828 | metrics[i] = {{"auth.meta_load", load.auth.meta_load()}, | |
829 | {"all.meta_load", load.all.meta_load()}, | |
830 | {"req_rate", load.req_rate}, | |
831 | {"queue_len", load.queue_len}, | |
832 | {"cpu_load_avg", load.cpu_load_avg}}; | |
833 | } | |
834 | ||
835 | /* execute the balancer */ | |
836 | Mantle mantle; | |
837 | int ret = mantle.balance(bal_code, mds->get_nodeid(), metrics, state.targets); | |
9f95a23c | 838 | dout(7) << " mantle decided that new targets=" << state.targets << dendl; |
7c673cae FG |
839 | |
840 | /* mantle doesn't know about cluster size, so check target len here */ | |
841 | if ((int) state.targets.size() != cluster_size) | |
842 | return -EINVAL; | |
843 | else if (ret) | |
844 | return ret; | |
845 | ||
846 | try_rebalance(state); | |
847 | return 0; | |
848 | } | |
849 | ||
850 | ||
851 | ||
852 | void MDBalancer::try_rebalance(balance_state_t& state) | |
853 | { | |
11fdf7f2 | 854 | if (g_conf()->mds_thrash_exports) { |
7c673cae FG |
855 | dout(5) << "mds_thrash is on; not performing standard rebalance operation!" |
856 | << dendl; | |
857 | return; | |
858 | } | |
859 | ||
860 | // make a sorted list of my imports | |
28e407b8 AA |
861 | multimap<double, CDir*> import_pop_map; |
862 | multimap<mds_rank_t, pair<CDir*, double> > import_from_map; | |
7c673cae | 863 | |
11fdf7f2 | 864 | for (auto& dir : mds->mdcache->get_fullauth_subtrees()) { |
28e407b8 AA |
865 | CInode *diri = dir->get_inode(); |
866 | if (diri->is_mdsdir()) | |
867 | continue; | |
868 | if (diri->get_export_pin(false) != MDS_RANK_NONE) | |
869 | continue; | |
870 | if (dir->is_freezing() || dir->is_frozen()) | |
871 | continue; // export pbly already in progress | |
7c673cae | 872 | |
28e407b8 | 873 | mds_rank_t from = diri->authority().first; |
11fdf7f2 TL |
874 | double pop = dir->pop_auth_subtree.meta_load(); |
875 | if (g_conf()->mds_bal_idle_threshold > 0 && | |
876 | pop < g_conf()->mds_bal_idle_threshold && | |
28e407b8 AA |
877 | diri != mds->mdcache->get_root() && |
878 | from != mds->get_nodeid()) { | |
879 | dout(5) << " exporting idle (" << pop << ") import " << *dir | |
880 | << " back to mds." << from << dendl; | |
881 | mds->mdcache->migrator->export_dir_nicely(dir, from); | |
7c673cae FG |
882 | continue; |
883 | } | |
884 | ||
28e407b8 AA |
885 | dout(15) << " map: i imported " << *dir << " from " << from << dendl; |
886 | import_pop_map.insert(make_pair(pop, dir)); | |
887 | import_from_map.insert(make_pair(from, make_pair(dir, pop))); | |
7c673cae FG |
888 | } |
889 | ||
7c673cae | 890 | // do my exports! |
28e407b8 | 891 | map<mds_rank_t, double> export_pop_map; |
7c673cae FG |
892 | |
893 | for (auto &it : state.targets) { | |
894 | mds_rank_t target = it.first; | |
895 | double amount = it.second; | |
896 | ||
28e407b8 AA |
897 | if (amount < MIN_OFFLOAD) |
898 | continue; | |
91327a77 AA |
899 | if (amount * 10 * state.targets.size() < target_load) |
900 | continue; | |
7c673cae FG |
901 | |
902 | dout(5) << "want to send " << amount << " to mds." << target | |
903 | //<< " .. " << (*it).second << " * " << load_fac | |
904 | << " -> " << amount | |
905 | << dendl;//" .. fudge is " << fudge << dendl; | |
7c673cae | 906 | |
28e407b8 | 907 | double& have = export_pop_map[target]; |
7c673cae FG |
908 | |
909 | mds->mdcache->show_subtrees(); | |
910 | ||
911 | // search imports from target | |
912 | if (import_from_map.count(target)) { | |
9f95a23c | 913 | dout(7) << " aha, looking through imports from target mds." << target << dendl; |
28e407b8 AA |
914 | for (auto p = import_from_map.equal_range(target); |
915 | p.first != p.second; ) { | |
916 | CDir *dir = p.first->second.first; | |
917 | double pop = p.first->second.second; | |
9f95a23c | 918 | dout(7) << "considering " << *dir << " from " << (*p.first).first << dendl; |
28e407b8 | 919 | auto plast = p.first++; |
7c673cae | 920 | |
28e407b8 | 921 | if (dir->inode->is_base()) |
7c673cae | 922 | continue; |
11fdf7f2 | 923 | ceph_assert(dir->inode->authority().first == target); // cuz that's how i put it in the map, dummy |
7c673cae FG |
924 | |
925 | if (pop <= amount-have) { | |
9f95a23c | 926 | dout(7) << "reexporting " << *dir << " pop " << pop |
7c673cae FG |
927 | << " back to mds." << target << dendl; |
928 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
929 | have += pop; | |
930 | import_from_map.erase(plast); | |
28e407b8 AA |
931 | for (auto q = import_pop_map.equal_range(pop); |
932 | q.first != q.second; ) { | |
933 | if (q.first->second == dir) { | |
934 | import_pop_map.erase(q.first); | |
935 | break; | |
936 | } | |
937 | q.first++; | |
938 | } | |
7c673cae | 939 | } else { |
9f95a23c | 940 | dout(7) << "can't reexport " << *dir << ", too big " << pop << dendl; |
7c673cae | 941 | } |
28e407b8 AA |
942 | if (amount-have < MIN_OFFLOAD) |
943 | break; | |
7c673cae FG |
944 | } |
945 | } | |
28e407b8 AA |
946 | } |
947 | ||
948 | // any other imports | |
949 | for (auto &it : state.targets) { | |
950 | mds_rank_t target = it.first; | |
951 | double amount = it.second; | |
952 | ||
953 | if (!export_pop_map.count(target)) | |
954 | continue; | |
955 | double& have = export_pop_map[target]; | |
956 | if (amount-have < MIN_OFFLOAD) | |
7c673cae | 957 | continue; |
7c673cae | 958 | |
28e407b8 AA |
959 | for (auto p = import_pop_map.begin(); |
960 | p != import_pop_map.end(); ) { | |
961 | CDir *dir = p->second; | |
962 | if (dir->inode->is_base()) { | |
963 | ++p; | |
964 | continue; | |
965 | } | |
7c673cae | 966 | |
28e407b8 AA |
967 | double pop = p->first; |
968 | if (pop <= amount-have && pop > MIN_REEXPORT) { | |
9f95a23c | 969 | dout(5) << "reexporting " << *dir << " pop " << pop |
28e407b8 AA |
970 | << " to mds." << target << dendl; |
971 | have += pop; | |
972 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
973 | import_pop_map.erase(p++); | |
974 | } else { | |
975 | ++p; | |
7c673cae | 976 | } |
28e407b8 AA |
977 | if (amount-have < MIN_OFFLOAD) |
978 | break; | |
7c673cae | 979 | } |
28e407b8 | 980 | } |
7c673cae | 981 | |
28e407b8 AA |
982 | set<CDir*> already_exporting; |
983 | ||
984 | for (auto &it : state.targets) { | |
985 | mds_rank_t target = it.first; | |
986 | double amount = it.second; | |
7c673cae | 987 | |
28e407b8 AA |
988 | if (!export_pop_map.count(target)) |
989 | continue; | |
990 | double& have = export_pop_map[target]; | |
991 | if (amount-have < MIN_OFFLOAD) | |
992 | continue; | |
993 | ||
994 | // okay, search for fragments of my workload | |
9f95a23c | 995 | std::vector<CDir*> exports; |
7c673cae | 996 | |
28e407b8 AA |
997 | for (auto p = import_pop_map.rbegin(); |
998 | p != import_pop_map.rend(); | |
999 | ++p) { | |
1000 | CDir *dir = p->second; | |
9f95a23c | 1001 | find_exports(dir, amount, &exports, have, already_exporting); |
28e407b8 | 1002 | if (amount-have < MIN_OFFLOAD) |
7c673cae FG |
1003 | break; |
1004 | } | |
1005 | //fudge = amount - have; | |
1006 | ||
9f95a23c | 1007 | for (const auto& dir : exports) { |
28e407b8 | 1008 | dout(5) << " - exporting " << dir->pop_auth_subtree |
11fdf7f2 | 1009 | << " " << dir->pop_auth_subtree.meta_load() |
28e407b8 AA |
1010 | << " to mds." << target << " " << *dir << dendl; |
1011 | mds->mdcache->migrator->export_dir_nicely(dir, target); | |
7c673cae FG |
1012 | } |
1013 | } | |
1014 | ||
9f95a23c | 1015 | dout(7) << "done" << dendl; |
7c673cae FG |
1016 | mds->mdcache->show_subtrees(); |
1017 | } | |
1018 | ||
7c673cae FG |
1019 | void MDBalancer::find_exports(CDir *dir, |
1020 | double amount, | |
9f95a23c | 1021 | std::vector<CDir*>* exports, |
7c673cae FG |
1022 | double& have, |
1023 | set<CDir*>& already_exporting) | |
1024 | { | |
11fdf7f2 TL |
1025 | auto now = clock::now(); |
1026 | auto duration = std::chrono::duration<double>(now-rebalance_time).count(); | |
1027 | if (duration > 0.1) { | |
28e407b8 AA |
1028 | derr << " balancer runs too long" << dendl_impl; |
1029 | have = amount; | |
1030 | return; | |
1031 | } | |
1032 | ||
11fdf7f2 | 1033 | ceph_assert(dir->is_auth()); |
28e407b8 | 1034 | |
7c673cae | 1035 | double need = amount - have; |
11fdf7f2 | 1036 | if (need < amount * g_conf()->mds_bal_min_start) |
7c673cae | 1037 | return; // good enough! |
28e407b8 | 1038 | |
11fdf7f2 TL |
1039 | double needmax = need * g_conf()->mds_bal_need_max; |
1040 | double needmin = need * g_conf()->mds_bal_need_min; | |
1041 | double midchunk = need * g_conf()->mds_bal_midchunk; | |
1042 | double minchunk = need * g_conf()->mds_bal_minchunk; | |
7c673cae | 1043 | |
9f95a23c | 1044 | std::vector<CDir*> bigger_rep, bigger_unrep; |
7c673cae FG |
1045 | multimap<double, CDir*> smaller; |
1046 | ||
11fdf7f2 | 1047 | double dir_pop = dir->pop_auth_subtree.meta_load(); |
9f95a23c | 1048 | dout(7) << "in " << dir_pop << " " << *dir << " need " << need << " (" << needmin << " - " << needmax << ")" << dendl; |
7c673cae FG |
1049 | |
1050 | double subdir_sum = 0; | |
28e407b8 AA |
1051 | for (elist<CInode*>::iterator it = dir->pop_lru_subdirs.begin_use_current(); |
1052 | !it.end(); ) { | |
1053 | CInode *in = *it; | |
1054 | ++it; | |
1055 | ||
11fdf7f2 TL |
1056 | ceph_assert(in->is_dir()); |
1057 | ceph_assert(in->get_parent_dir() == dir); | |
7c673cae | 1058 | |
9f95a23c | 1059 | auto&& dfls = in->get_nested_dirfrags(); |
28e407b8 AA |
1060 | |
1061 | size_t num_idle_frags = 0; | |
9f95a23c | 1062 | for (const auto& subdir : dfls) { |
28e407b8 AA |
1063 | if (already_exporting.count(subdir)) |
1064 | continue; | |
7c673cae | 1065 | |
28e407b8 AA |
1066 | // we know all ancestor dirfrags up to subtree root are not freezing or frozen. |
1067 | // It's more efficient to use CDir::is_{freezing,frozen}_tree_root() | |
1068 | if (subdir->is_frozen_dir() || subdir->is_frozen_tree_root() || | |
1069 | subdir->is_freezing_dir() || subdir->is_freezing_tree_root()) | |
1070 | continue; // can't export this right now! | |
7c673cae FG |
1071 | |
1072 | // how popular? | |
11fdf7f2 | 1073 | double pop = subdir->pop_auth_subtree.meta_load(); |
7c673cae FG |
1074 | subdir_sum += pop; |
1075 | dout(15) << " subdir pop " << pop << " " << *subdir << dendl; | |
1076 | ||
28e407b8 AA |
1077 | if (pop < minchunk) { |
1078 | num_idle_frags++; | |
1079 | continue; | |
1080 | } | |
7c673cae FG |
1081 | |
1082 | // lucky find? | |
1083 | if (pop > needmin && pop < needmax) { | |
9f95a23c | 1084 | exports->push_back(subdir); |
7c673cae FG |
1085 | already_exporting.insert(subdir); |
1086 | have += pop; | |
1087 | return; | |
1088 | } | |
1089 | ||
1090 | if (pop > need) { | |
1091 | if (subdir->is_rep()) | |
1092 | bigger_rep.push_back(subdir); | |
1093 | else | |
1094 | bigger_unrep.push_back(subdir); | |
1095 | } else | |
1096 | smaller.insert(pair<double,CDir*>(pop, subdir)); | |
1097 | } | |
28e407b8 AA |
1098 | if (dfls.size() == num_idle_frags) |
1099 | in->item_pop_lru.remove_myself(); | |
7c673cae FG |
1100 | } |
1101 | dout(15) << " sum " << subdir_sum << " / " << dir_pop << dendl; | |
1102 | ||
1103 | // grab some sufficiently big small items | |
1104 | multimap<double,CDir*>::reverse_iterator it; | |
1105 | for (it = smaller.rbegin(); | |
1106 | it != smaller.rend(); | |
1107 | ++it) { | |
1108 | ||
1109 | if ((*it).first < midchunk) | |
1110 | break; // try later | |
1111 | ||
1112 | dout(7) << " taking smaller " << *(*it).second << dendl; | |
1113 | ||
9f95a23c | 1114 | exports->push_back((*it).second); |
7c673cae FG |
1115 | already_exporting.insert((*it).second); |
1116 | have += (*it).first; | |
1117 | if (have > needmin) | |
1118 | return; | |
1119 | } | |
1120 | ||
1121 | // apprently not enough; drill deeper into the hierarchy (if non-replicated) | |
9f95a23c TL |
1122 | for (const auto& dir : bigger_unrep) { |
1123 | dout(15) << " descending into " << *dir << dendl; | |
1124 | find_exports(dir, amount, exports, have, already_exporting); | |
7c673cae FG |
1125 | if (have > needmin) |
1126 | return; | |
1127 | } | |
1128 | ||
1129 | // ok fine, use smaller bits | |
1130 | for (; | |
1131 | it != smaller.rend(); | |
1132 | ++it) { | |
1133 | dout(7) << " taking (much) smaller " << it->first << " " << *(*it).second << dendl; | |
1134 | ||
9f95a23c | 1135 | exports->push_back((*it).second); |
7c673cae FG |
1136 | already_exporting.insert((*it).second); |
1137 | have += (*it).first; | |
1138 | if (have > needmin) | |
1139 | return; | |
1140 | } | |
1141 | ||
1142 | // ok fine, drill into replicated dirs | |
9f95a23c TL |
1143 | for (const auto& dir : bigger_rep) { |
1144 | dout(7) << " descending into replicated " << *dir << dendl; | |
1145 | find_exports(dir, amount, exports, have, already_exporting); | |
7c673cae FG |
1146 | if (have > needmin) |
1147 | return; | |
1148 | } | |
7c673cae FG |
1149 | } |
1150 | ||
11fdf7f2 | 1151 | void MDBalancer::hit_inode(CInode *in, int type, int who) |
7c673cae FG |
1152 | { |
1153 | // hit inode | |
11fdf7f2 | 1154 | in->pop.get(type).hit(); |
7c673cae FG |
1155 | |
1156 | if (in->get_parent_dn()) | |
11fdf7f2 | 1157 | hit_dir(in->get_parent_dn()->get_dir(), type, who); |
7c673cae FG |
1158 | } |
1159 | ||
1160 | void MDBalancer::maybe_fragment(CDir *dir, bool hot) | |
1161 | { | |
1162 | // split/merge | |
11fdf7f2 | 1163 | if (bal_fragment_dirs && bal_fragment_interval > 0 && |
91327a77 | 1164 | dir->is_auth() && |
11fdf7f2 | 1165 | !dir->inode->is_base() && // not root/mdsdir (for now at least) |
91327a77 | 1166 | !dir->inode->is_stray()) { // not straydir |
7c673cae FG |
1167 | |
1168 | // split | |
11fdf7f2 | 1169 | if (g_conf()->mds_bal_split_size > 0 && (dir->should_split() || hot)) { |
7c673cae FG |
1170 | if (split_pending.count(dir->dirfrag()) == 0) { |
1171 | queue_split(dir, false); | |
1172 | } else { | |
1173 | if (dir->should_split_fast()) { | |
1174 | queue_split(dir, true); | |
1175 | } else { | |
9f95a23c | 1176 | dout(10) << ": fragment already enqueued to split: " |
7c673cae FG |
1177 | << *dir << dendl; |
1178 | } | |
1179 | } | |
1180 | } | |
1181 | ||
1182 | // merge? | |
1183 | if (dir->get_frag() != frag_t() && dir->should_merge() && | |
1184 | merge_pending.count(dir->dirfrag()) == 0) { | |
1185 | queue_merge(dir); | |
1186 | } | |
1187 | } | |
1188 | } | |
1189 | ||
11fdf7f2 | 1190 | void MDBalancer::hit_dir(CDir *dir, int type, int who, double amount) |
7c673cae FG |
1191 | { |
1192 | // hit me | |
11fdf7f2 | 1193 | double v = dir->pop_me.get(type).hit(amount); |
7c673cae | 1194 | |
11fdf7f2 TL |
1195 | const bool hot = (v > g_conf()->mds_bal_split_rd && type == META_POP_IRD) || |
1196 | (v > g_conf()->mds_bal_split_wr && type == META_POP_IWR); | |
7c673cae | 1197 | |
9f95a23c | 1198 | dout(20) << type << " pop is " << v << ", frag " << dir->get_frag() |
11fdf7f2 | 1199 | << " size " << dir->get_frag_size() << " " << dir->pop_me << dendl; |
7c673cae FG |
1200 | |
1201 | maybe_fragment(dir, hot); | |
1202 | ||
1203 | // replicate? | |
1204 | if (type == META_POP_IRD && who >= 0) { | |
11fdf7f2 | 1205 | dir->pop_spread.hit(who); |
7c673cae FG |
1206 | } |
1207 | ||
1208 | double rd_adj = 0.0; | |
1209 | if (type == META_POP_IRD && | |
1210 | dir->last_popularity_sample < last_sample) { | |
11fdf7f2 | 1211 | double dir_pop = dir->pop_auth_subtree.get(type).get(); // hmm?? |
7c673cae | 1212 | dir->last_popularity_sample = last_sample; |
11fdf7f2 | 1213 | double pop_sp = dir->pop_spread.get(); |
7c673cae FG |
1214 | dir_pop += pop_sp * 10; |
1215 | ||
1216 | //if (dir->ino() == inodeno_t(0x10000000002)) | |
1217 | if (pop_sp > 0) { | |
9f95a23c | 1218 | dout(20) << type << " pop " << dir_pop << " spread " << pop_sp |
7c673cae FG |
1219 | << " " << dir->pop_spread.last[0] |
1220 | << " " << dir->pop_spread.last[1] | |
1221 | << " " << dir->pop_spread.last[2] | |
1222 | << " " << dir->pop_spread.last[3] | |
1223 | << " in " << *dir << dendl; | |
1224 | } | |
1225 | ||
1226 | if (dir->is_auth() && !dir->is_ambiguous_auth()) { | |
1227 | if (!dir->is_rep() && | |
11fdf7f2 | 1228 | dir_pop >= g_conf()->mds_bal_replicate_threshold) { |
7c673cae | 1229 | // replicate |
11fdf7f2 | 1230 | double rdp = dir->pop_me.get(META_POP_IRD).get(); |
7c673cae FG |
1231 | rd_adj = rdp / mds->get_mds_map()->get_num_in_mds() - rdp; |
1232 | rd_adj /= 2.0; // temper somewhat | |
1233 | ||
b32b8144 | 1234 | dout(5) << "replicating dir " << *dir << " pop " << dir_pop << " .. rdp " << rdp << " adj " << rd_adj << dendl; |
7c673cae FG |
1235 | |
1236 | dir->dir_rep = CDir::REP_ALL; | |
1237 | mds->mdcache->send_dir_updates(dir, true); | |
1238 | ||
1239 | // fixme this should adjust the whole pop hierarchy | |
1240 | dir->pop_me.get(META_POP_IRD).adjust(rd_adj); | |
1241 | dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); | |
1242 | } | |
1243 | ||
1244 | if (dir->ino() != 1 && | |
1245 | dir->is_rep() && | |
11fdf7f2 | 1246 | dir_pop < g_conf()->mds_bal_unreplicate_threshold) { |
7c673cae | 1247 | // unreplicate |
b32b8144 | 1248 | dout(5) << "unreplicating dir " << *dir << " pop " << dir_pop << dendl; |
7c673cae FG |
1249 | |
1250 | dir->dir_rep = CDir::REP_NONE; | |
1251 | mds->mdcache->send_dir_updates(dir); | |
1252 | } | |
1253 | } | |
1254 | } | |
1255 | ||
1256 | // adjust ancestors | |
1257 | bool hit_subtree = dir->is_auth(); // current auth subtree (if any) | |
1258 | bool hit_subtree_nested = dir->is_auth(); // all nested auth subtrees | |
1259 | ||
1260 | while (true) { | |
28e407b8 | 1261 | CDir *pdir = dir->inode->get_parent_dir(); |
11fdf7f2 | 1262 | dir->pop_nested.get(type).hit(amount); |
7c673cae | 1263 | if (rd_adj != 0.0) |
11fdf7f2 | 1264 | dir->pop_nested.get(META_POP_IRD).adjust(rd_adj); |
7c673cae FG |
1265 | |
1266 | if (hit_subtree) { | |
11fdf7f2 | 1267 | dir->pop_auth_subtree.get(type).hit(amount); |
28e407b8 | 1268 | |
7c673cae | 1269 | if (rd_adj != 0.0) |
11fdf7f2 | 1270 | dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); |
28e407b8 AA |
1271 | |
1272 | if (dir->is_subtree_root()) | |
1273 | hit_subtree = false; // end of auth domain, stop hitting auth counters. | |
1274 | else if (pdir) | |
1275 | pdir->pop_lru_subdirs.push_front(&dir->get_inode()->item_pop_lru); | |
7c673cae FG |
1276 | } |
1277 | ||
1278 | if (hit_subtree_nested) { | |
11fdf7f2 | 1279 | dir->pop_auth_subtree_nested.get(type).hit(amount); |
7c673cae | 1280 | if (rd_adj != 0.0) |
11fdf7f2 | 1281 | dir->pop_auth_subtree_nested.get(META_POP_IRD).adjust(rd_adj); |
7c673cae | 1282 | } |
28e407b8 AA |
1283 | if (!pdir) break; |
1284 | dir = pdir; | |
7c673cae FG |
1285 | } |
1286 | } | |
1287 | ||
1288 | ||
1289 | /* | |
1290 | * subtract off an exported chunk. | |
1291 | * this excludes *dir itself (encode_export_dir should have take care of that) | |
1292 | * we _just_ do the parents' nested counters. | |
1293 | * | |
1294 | * NOTE: call me _after_ forcing *dir into a subtree root, | |
1295 | * but _before_ doing the encode_export_dirs. | |
1296 | */ | |
11fdf7f2 | 1297 | void MDBalancer::subtract_export(CDir *dir) |
7c673cae FG |
1298 | { |
1299 | dirfrag_load_vec_t subload = dir->pop_auth_subtree; | |
1300 | ||
1301 | while (true) { | |
1302 | dir = dir->inode->get_parent_dir(); | |
1303 | if (!dir) break; | |
1304 | ||
11fdf7f2 TL |
1305 | dir->pop_nested.sub(subload); |
1306 | dir->pop_auth_subtree_nested.sub(subload); | |
7c673cae FG |
1307 | } |
1308 | } | |
1309 | ||
1310 | ||
11fdf7f2 | 1311 | void MDBalancer::add_import(CDir *dir) |
7c673cae FG |
1312 | { |
1313 | dirfrag_load_vec_t subload = dir->pop_auth_subtree; | |
1314 | ||
1315 | while (true) { | |
1316 | dir = dir->inode->get_parent_dir(); | |
1317 | if (!dir) break; | |
1318 | ||
11fdf7f2 TL |
1319 | dir->pop_nested.add(subload); |
1320 | dir->pop_auth_subtree_nested.add(subload); | |
7c673cae FG |
1321 | } |
1322 | } | |
1323 | ||
11fdf7f2 | 1324 | void MDBalancer::adjust_pop_for_rename(CDir *pdir, CDir *dir, bool inc) |
28e407b8 | 1325 | { |
28e407b8 AA |
1326 | bool adjust_subtree_nest = dir->is_auth(); |
1327 | bool adjust_subtree = adjust_subtree_nest && !dir->is_subtree_root(); | |
1328 | CDir *cur = dir; | |
1329 | while (true) { | |
1330 | if (inc) { | |
11fdf7f2 | 1331 | pdir->pop_nested.add(dir->pop_nested); |
28e407b8 | 1332 | if (adjust_subtree) { |
11fdf7f2 | 1333 | pdir->pop_auth_subtree.add(dir->pop_auth_subtree); |
28e407b8 AA |
1334 | pdir->pop_lru_subdirs.push_front(&cur->get_inode()->item_pop_lru); |
1335 | } | |
1336 | ||
1337 | if (adjust_subtree_nest) | |
11fdf7f2 | 1338 | pdir->pop_auth_subtree_nested.add(dir->pop_auth_subtree_nested); |
28e407b8 | 1339 | } else { |
11fdf7f2 | 1340 | pdir->pop_nested.sub(dir->pop_nested); |
28e407b8 | 1341 | if (adjust_subtree) |
11fdf7f2 | 1342 | pdir->pop_auth_subtree.sub(dir->pop_auth_subtree); |
28e407b8 AA |
1343 | |
1344 | if (adjust_subtree_nest) | |
11fdf7f2 | 1345 | pdir->pop_auth_subtree_nested.sub(dir->pop_auth_subtree_nested); |
28e407b8 AA |
1346 | } |
1347 | ||
1348 | if (pdir->is_subtree_root()) | |
1349 | adjust_subtree = false; | |
1350 | cur = pdir; | |
1351 | pdir = pdir->inode->get_parent_dir(); | |
1352 | if (!pdir) break; | |
1353 | } | |
1354 | } | |
1355 | ||
224ce89b WB |
1356 | void MDBalancer::handle_mds_failure(mds_rank_t who) |
1357 | { | |
1358 | if (0 == who) { | |
28e407b8 | 1359 | mds_last_epoch_under_map.clear(); |
224ce89b WB |
1360 | } |
1361 | } | |
28e407b8 | 1362 | |
9f95a23c | 1363 | int MDBalancer::dump_loads(Formatter *f) const |
28e407b8 | 1364 | { |
9f95a23c | 1365 | std::deque<CDir*> dfs; |
28e407b8 AA |
1366 | if (mds->mdcache->get_root()) { |
1367 | mds->mdcache->get_root()->get_dirfrags(dfs); | |
1368 | } else { | |
9f95a23c | 1369 | dout(10) << "no root" << dendl; |
28e407b8 AA |
1370 | } |
1371 | ||
1372 | f->open_object_section("loads"); | |
1373 | ||
1374 | f->open_array_section("dirfrags"); | |
1375 | while (!dfs.empty()) { | |
1376 | CDir *dir = dfs.front(); | |
1377 | dfs.pop_front(); | |
1378 | ||
11fdf7f2 TL |
1379 | f->open_object_section("dir"); |
1380 | dir->dump_load(f); | |
1381 | f->close_section(); | |
28e407b8 AA |
1382 | |
1383 | for (auto it = dir->begin(); it != dir->end(); ++it) { | |
1384 | CInode *in = it->second->get_linkage()->get_inode(); | |
1385 | if (!in || !in->is_dir()) | |
1386 | continue; | |
1387 | ||
9f95a23c TL |
1388 | auto&& ls = in->get_dirfrags(); |
1389 | for (const auto& subdir : ls) { | |
28e407b8 AA |
1390 | if (subdir->pop_nested.meta_load() < .001) |
1391 | continue; | |
1392 | dfs.push_back(subdir); | |
1393 | } | |
1394 | } | |
1395 | } | |
1396 | f->close_section(); // dirfrags array | |
1397 | ||
1398 | f->open_object_section("mds_load"); | |
1399 | { | |
1400 | ||
11fdf7f2 | 1401 | auto dump_mds_load = [f](mds_load_t& load) { |
28e407b8 AA |
1402 | f->dump_float("request_rate", load.req_rate); |
1403 | f->dump_float("cache_hit_rate", load.cache_hit_rate); | |
1404 | f->dump_float("queue_length", load.queue_len); | |
1405 | f->dump_float("cpu_load", load.cpu_load_avg); | |
1406 | f->dump_float("mds_load", load.mds_load()); | |
1407 | ||
28e407b8 | 1408 | f->open_object_section("auth_dirfrags"); |
11fdf7f2 | 1409 | load.auth.dump(f); |
28e407b8 AA |
1410 | f->close_section(); |
1411 | f->open_object_section("all_dirfrags"); | |
11fdf7f2 | 1412 | load.all.dump(f); |
28e407b8 AA |
1413 | f->close_section(); |
1414 | }; | |
1415 | ||
1416 | for (auto p : mds_load) { | |
1417 | stringstream name; | |
1418 | name << "mds." << p.first; | |
1419 | f->open_object_section(name.str().c_str()); | |
1420 | dump_mds_load(p.second); | |
1421 | f->close_section(); | |
1422 | } | |
1423 | } | |
1424 | f->close_section(); // mds_load | |
1425 | ||
1426 | f->open_object_section("mds_meta_load"); | |
1427 | for (auto p : mds_meta_load) { | |
1428 | stringstream name; | |
1429 | name << "mds." << p.first; | |
1430 | f->dump_float(name.str().c_str(), p.second); | |
1431 | } | |
1432 | f->close_section(); // mds_meta_load | |
1433 | ||
1434 | f->open_object_section("mds_import_map"); | |
1435 | for (auto p : mds_import_map) { | |
1436 | stringstream name1; | |
1437 | name1 << "mds." << p.first; | |
1438 | f->open_array_section(name1.str().c_str()); | |
1439 | for (auto q : p.second) { | |
1440 | f->open_object_section("from"); | |
1441 | stringstream name2; | |
1442 | name2 << "mds." << q.first; | |
1443 | f->dump_float(name2.str().c_str(), q.second); | |
1444 | f->close_section(); | |
1445 | } | |
1446 | f->close_section(); // mds.? array | |
1447 | } | |
1448 | f->close_section(); // mds_import_map | |
1449 | ||
1450 | f->close_section(); // loads | |
1451 | return 0; | |
1452 | } |