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