-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathMDBalancer.cc
1544 lines (1300 loc) · 45.9 KB
/
MDBalancer.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
// vim: ts=8 sw=2 smarttab
/*
* Ceph - scalable distributed file system
*
* Copyright (C) 2004-2006 Sage Weil <[email protected]>
*
* This is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License version 2.1, as published by the Free Software
* Foundation. See file COPYING.
*
*/
#include "include/compat.h"
#include "mdstypes.h"
#include "mon/MonClient.h"
#include "MDBalancer.h"
#include "MDSRank.h"
#include "MDSMap.h"
#include "CInode.h"
#include "CDir.h"
#include "MDCache.h"
#include "Migrator.h"
#include "Mantle.h"
#include "include/Context.h"
#include "msg/Messenger.h"
#include <fstream>
#include <vector>
#include <map>
using namespace std;
#include "common/config.h"
#include "common/errno.h"
/* Note, by default debug_mds_balancer is 1/5. For debug messages 1<lvl<=5,
* should_gather (below) will be true; so, debug_mds will be ignored even if
* set to 20/20. For this reason, some messages may not appear in the log.
* Increase both debug levels to get expected output!
*/
#define dout_context g_ceph_context
#undef dout_prefix
#define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".bal " << __func__ << " "
#undef dout
#define dout(lvl) \
do {\
auto subsys = ceph_subsys_mds;\
if ((dout_context)->_conf->subsys.should_gather(ceph_subsys_mds_balancer, lvl)) {\
subsys = ceph_subsys_mds_balancer;\
}\
dout_impl(dout_context, ceph::dout::need_dynamic(subsys), lvl) dout_prefix
#undef dendl
#define dendl dendl_impl; } while (0)
#define MIN_LOAD 50 // ??
#define MIN_REEXPORT 5 // will automatically reexport
#define MIN_OFFLOAD 10 // point at which i stop trying, close enough
int MDBalancer::proc_message(const cref_t<Message> &m)
{
switch (m->get_type()) {
case MSG_MDS_HEARTBEAT:
handle_heartbeat(ref_cast<MHeartbeat>(m));
break;
default:
derr << " balancer unknown message " << m->get_type() << dendl_impl;
ceph_abort_msg("balancer unknown message");
}
return 0;
}
MDBalancer::MDBalancer(MDSRank *m, Messenger *msgr, MonClient *monc) :
mds(m), messenger(msgr), mon_client(monc)
{
bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs");
bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval");
}
void MDBalancer::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mds_map)
{
if (changed.count("mds_bal_fragment_dirs")) {
bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs");
}
if (changed.count("mds_bal_fragment_interval")) {
bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval");
}
}
bool MDBalancer::test_rank_mask(mds_rank_t rank)
{
return mds->mdsmap->get_bal_rank_mask_bitset().test(rank);
}
void MDBalancer::handle_export_pins(void)
{
const mds_rank_t max_mds = mds->mdsmap->get_max_mds();
auto mdcache = mds->mdcache;
auto &q = mdcache->export_pin_queue;
auto it = q.begin();
dout(20) << "export_pin_queue size=" << q.size() << dendl;
while (it != q.end()) {
auto cur = it++;
CInode *in = *cur;
ceph_assert(in->is_dir());
mds_rank_t export_pin = in->get_export_pin(false);
in->check_pin_policy(export_pin);
if (export_pin >= max_mds) {
dout(20) << " delay export_pin=" << export_pin << " on " << *in << dendl;
in->state_clear(CInode::STATE_QUEUEDEXPORTPIN);
q.erase(cur);
in->state_set(CInode::STATE_DELAYEDEXPORTPIN);
mdcache->export_pin_delayed_queue.insert(in);
continue;
}
dout(20) << " executing export_pin=" << export_pin << " on " << *in << dendl;
unsigned min_frag_bits = 0;
mds_rank_t target = MDS_RANK_NONE;
if (export_pin >= 0)
target = export_pin;
else if (export_pin == MDS_RANK_EPHEMERAL_RAND)
target = mdcache->hash_into_rank_bucket(in->ino());
else if (export_pin == MDS_RANK_EPHEMERAL_DIST)
min_frag_bits = mdcache->get_ephemeral_dist_frag_bits();
bool remove = true;
for (auto&& dir : in->get_dirfrags()) {
if (!dir->is_auth())
continue;
if (export_pin == MDS_RANK_EPHEMERAL_DIST) {
if (dir->get_frag().bits() < min_frag_bits) {
if (!dir->state_test(CDir::STATE_CREATING) &&
!dir->is_frozen() && !dir->is_freezing()) {
queue_split(dir, true);
}
remove = false;
continue;
}
target = mdcache->hash_into_rank_bucket(in->ino(), dir->get_frag());
}
if (target == MDS_RANK_NONE) {
if (dir->state_test(CDir::STATE_AUXSUBTREE)) {
if (dir->is_frozen() || dir->is_freezing()) {
// try again later
remove = false;
continue;
}
dout(10) << " clear auxsubtree on " << *dir << dendl;
dir->state_clear(CDir::STATE_AUXSUBTREE);
mds->mdcache->try_subtree_merge(dir);
}
} else if (target == mds->get_nodeid()) {
if (dir->state_test(CDir::STATE_AUXSUBTREE)) {
ceph_assert(dir->is_subtree_root());
} else if (dir->state_test(CDir::STATE_CREATING) ||
dir->is_frozen() || dir->is_freezing()) {
// try again later
remove = false;
continue;
} else if (!dir->is_subtree_root()) {
dir->state_set(CDir::STATE_AUXSUBTREE);
mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid());
dout(10) << " create aux subtree on " << *dir << dendl;
} else {
dout(10) << " set auxsubtree bit on " << *dir << dendl;
dir->state_set(CDir::STATE_AUXSUBTREE);
}
} else {
/* Only export a directory if it's non-empty. An empty directory will
* be sent back by the importer.
*/
if (dir->get_num_head_items() > 0) {
mds->mdcache->migrator->export_dir(dir, target);
}
remove = false;
}
}
if (remove) {
in->state_clear(CInode::STATE_QUEUEDEXPORTPIN);
q.erase(cur);
}
}
std::vector<CDir*> authsubs = mdcache->get_auth_subtrees();
bool print_auth_subtrees = true;
if (authsubs.size() > AUTH_TREES_THRESHOLD &&
!g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) {
dout(15) << "number of auth trees = " << authsubs.size() << "; not "
"printing auth trees" << dendl;
print_auth_subtrees = false;
}
for (auto &cd : authsubs) {
mds_rank_t export_pin = cd->inode->get_export_pin();
cd->inode->check_pin_policy(export_pin);
if (export_pin == MDS_RANK_EPHEMERAL_DIST) {
export_pin = mdcache->hash_into_rank_bucket(cd->ino(), cd->get_frag());
} else if (export_pin == MDS_RANK_EPHEMERAL_RAND) {
export_pin = mdcache->hash_into_rank_bucket(cd->ino());
}
if (print_auth_subtrees)
dout(25) << "auth tree " << *cd << " export_pin=" << export_pin << dendl;
if (export_pin >= 0 && export_pin != mds->get_nodeid() &&
export_pin < mds->mdsmap->get_max_mds()) {
mdcache->migrator->export_dir(cd, export_pin);
}
}
}
void MDBalancer::tick()
{
static int num_bal_times = g_conf()->mds_bal_max;
bool balance_automate = mds->mdsmap->allows_balance_automate();
auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval");
auto bal_max_until = g_conf().get_val<int64_t>("mds_bal_max_until");
time now = clock::now();
if (g_conf()->mds_bal_export_pin) {
handle_export_pins();
}
// sample?
if (chrono::duration<double>(now-last_sample).count() >
g_conf()->mds_bal_sample_interval) {
dout(15) << "tick last_sample now " << now << dendl;
last_sample = now;
}
// We can use duration_cast below, although the result is an int,
// because the values from g_conf are also integers.
// balance?
if (balance_automate
&& mds->get_nodeid() == 0
&& mds->is_active()
&& bal_interval > 0
&& chrono::duration_cast<chrono::seconds>(now - last_heartbeat).count() >= bal_interval
&& (num_bal_times || (bal_max_until >= 0 && mds->get_uptime().count() > bal_max_until))) {
last_heartbeat = now;
send_heartbeat();
num_bal_times--;
}
mds->mdcache->show_subtrees(10, true);
}
class C_Bal_SendHeartbeat : public MDSInternalContext {
public:
explicit C_Bal_SendHeartbeat(MDSRank *mds_) : MDSInternalContext(mds_) { }
void finish(int f) override {
mds->balancer->send_heartbeat();
}
};
double mds_load_t::mds_load() const
{
switch(g_conf()->mds_bal_mode) {
case 0:
return
.8 * auth.meta_load() +
.2 * all.meta_load() +
req_rate +
10.0 * queue_len;
case 1:
return req_rate + 10.0*queue_len;
case 2:
return cpu_load_avg;
}
ceph_abort();
return 0;
}
mds_load_t MDBalancer::get_load()
{
auto now = clock::now();
mds_load_t load{DecayRate()}; /* zero DecayRate! */
if (mds->mdcache->get_root()) {
auto&& ls = mds->mdcache->get_root()->get_dirfrags();
for (auto &d : ls) {
load.auth.add(d->pop_auth_subtree_nested);
load.all.add(d->pop_nested);
}
} else {
dout(20) << "no root, no load" << dendl;
}
uint64_t num_requests = mds->get_num_requests();
uint64_t num_traverse = mds->logger->get(l_mds_traverse);
uint64_t num_traverse_hit = mds->logger->get(l_mds_traverse_hit);
uint64_t cpu_time = 1;
{
string stat_path = PROCPREFIX "/proc/self/stat";
ifstream stat_file(stat_path);
if (stat_file.is_open()) {
vector<string> stat_vec(std::istream_iterator<string>{stat_file},
std::istream_iterator<string>());
if (stat_vec.size() >= 15) {
// utime + stime
cpu_time = strtoll(stat_vec[13].c_str(), nullptr, 10) +
strtoll(stat_vec[14].c_str(), nullptr, 10);
} else {
derr << "input file '" << stat_path << "' not resolvable" << dendl_impl;
}
} else {
derr << "input file '" << stat_path << "' not found" << dendl_impl;
}
}
load.queue_len = messenger->get_dispatch_queue_len();
bool update_last = true;
if (last_get_load != clock::zero() &&
now > last_get_load) {
double el = std::chrono::duration<double>(now-last_get_load).count();
if (el >= 1.0) {
if (num_requests > last_num_requests)
load.req_rate = (num_requests - last_num_requests) / el;
if (cpu_time > last_cpu_time)
load.cpu_load_avg = (cpu_time - last_cpu_time) / el;
if (num_traverse > last_num_traverse && num_traverse_hit > last_num_traverse_hit)
load.cache_hit_rate = (double)(num_traverse_hit - last_num_traverse_hit) / (num_traverse - last_num_traverse);
} else {
auto p = mds_load.find(mds->get_nodeid());
if (p != mds_load.end()) {
load.req_rate = p->second.req_rate;
load.cpu_load_avg = p->second.cpu_load_avg;
load.cache_hit_rate = p->second.cache_hit_rate;
}
if (num_requests >= last_num_requests && cpu_time >= last_cpu_time &&
num_traverse >= last_num_traverse && num_traverse_hit >= last_num_traverse_hit)
update_last = false;
}
}
if (update_last) {
last_num_requests = num_requests;
last_cpu_time = cpu_time;
last_get_load = now;
last_num_traverse = num_traverse;
last_num_traverse_hit = num_traverse_hit;
}
dout(15) << load << dendl;
return load;
}
/*
* Read synchronously from RADOS using a timeout. We cannot do daemon-local
* fallbacks (i.e. kick off async read when we are processing the map and
* check status when we get here) with the way the mds is structured.
*/
int MDBalancer::localize_balancer()
{
/* reset everything */
bool ack = false;
int r = 0;
bufferlist lua_src;
ceph::mutex lock = ceph::make_mutex("lock");
ceph::condition_variable cond;
/* we assume that balancer is in the metadata pool */
object_t oid = object_t(mds->mdsmap->get_balancer());
object_locator_t oloc(mds->get_metadata_pool());
ceph_tid_t tid = mds->objecter->read(oid, oloc, 0, 0, CEPH_NOSNAP, &lua_src, 0,
new C_SafeCond(lock, cond, &ack, &r));
dout(15) << "launched non-blocking read tid=" << tid
<< " oid=" << oid << " oloc=" << oloc << dendl;
/* timeout: if we waste half our time waiting for RADOS, then abort! */
std::cv_status ret_t = [&] {
auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval");
std::unique_lock locker{lock};
return cond.wait_for(locker, std::chrono::seconds(bal_interval / 2));
}();
/* success: store the balancer in memory and set the version. */
if (!r) {
if (ret_t == std::cv_status::timeout) {
mds->objecter->op_cancel(tid, -CEPHFS_ECANCELED);
return -CEPHFS_ETIMEDOUT;
}
bal_code.assign(lua_src.to_str());
bal_version.assign(oid.name);
dout(10) "bal_code=" << bal_code << dendl;
}
return r;
}
void MDBalancer::send_heartbeat()
{
if (mds->is_cluster_degraded()) {
dout(10) << "degraded" << dendl;
return;
}
if (!mds->mdcache->is_open()) {
dout(10) << "not open" << dendl;
mds->mdcache->wait_for_open(new C_Bal_SendHeartbeat(mds));
return;
}
if (mds->get_nodeid() == 0) {
beat_epoch++;
mds_load.clear();
}
// my load
mds_load_t load = get_load();
mds->logger->set(l_mds_load_cent, 100 * load.mds_load());
mds->logger->set(l_mds_dispatch_queue_len, load.queue_len);
auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(mds->get_nodeid()), std::forward_as_tuple(load));
if (!em.second) {
em.first->second = load;
}
// import_map -- how much do i import from whom
map<mds_rank_t, float> import_map;
for (auto& im : mds->mdcache->get_auth_subtrees()) {
mds_rank_t from = im->inode->authority().first;
if (from == mds->get_nodeid()) continue;
if (im->get_inode()->is_stray()) continue;
import_map[from] += im->pop_auth_subtree.meta_load();
}
mds_import_map[ mds->get_nodeid() ] = import_map;
dout(3) << " epoch " << beat_epoch << " load " << load << dendl;
for (const auto& [rank, load] : import_map) {
dout(5) << " import_map from " << rank << " -> " << load << dendl;
}
set<mds_rank_t> up;
mds->get_mds_map()->get_up_mds_set(up);
for (const auto& r : up) {
if (r == mds->get_nodeid())
continue;
auto hb = make_message<MHeartbeat>(load, beat_epoch);
hb->get_import_map() = import_map;
mds->send_message_mds(hb, r);
}
}
void MDBalancer::handle_heartbeat(const cref_t<MHeartbeat> &m)
{
mds_rank_t who = mds_rank_t(m->get_source().num());
dout(25) << "=== got heartbeat " << m->get_beat() << " from " << m->get_source().num() << " " << m->get_load() << dendl;
if (!mds->is_active())
return;
if (!mds->mdcache->is_open()) {
dout(10) << "opening root on handle_heartbeat" << dendl;
mds->mdcache->wait_for_open(new C_MDS_RetryMessage(mds, m));
return;
}
if (mds->is_cluster_degraded()) {
dout(10) << " degraded, ignoring" << dendl;
return;
}
if (mds->get_nodeid() != 0 && m->get_beat() > beat_epoch) {
dout(10) << "receive next epoch " << m->get_beat() << " from mds." << who << " before mds0" << dendl;
beat_epoch = m->get_beat();
// clear the mds load info whose epoch is less than beat_epoch
mds_load.clear();
}
if (who == 0) {
dout(20) << " from mds0, new epoch " << m->get_beat() << dendl;
if (beat_epoch != m->get_beat()) {
beat_epoch = m->get_beat();
mds_load.clear();
}
send_heartbeat();
mds->mdcache->show_subtrees();
} else if (mds->get_nodeid() == 0) {
if (beat_epoch != m->get_beat()) {
dout(10) << " old heartbeat epoch, ignoring" << dendl;
return;
}
}
{
auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(who), std::forward_as_tuple(m->get_load()));
if (!em.second) {
em.first->second = m->get_load();
}
}
mds_import_map[who] = m->get_import_map();
mds->mdsmap->update_num_mdss_in_rank_mask_bitset();
if (mds->mdsmap->get_num_mdss_in_rank_mask_bitset() > 0)
{
unsigned cluster_size = mds->get_mds_map()->get_num_in_mds();
if (mds_load.size() == cluster_size) {
// let's go!
//export_empties(); // no!
/* avoid spamming ceph -w if user does not turn mantle on */
if (mds->mdsmap->get_balancer() != "") {
int r = mantle_prep_rebalance();
if (!r) return;
mds->clog->warn() << "using old balancer; mantle failed for "
<< "balancer=" << mds->mdsmap->get_balancer()
<< " : " << cpp_strerror(r);
}
prep_rebalance(m->get_beat());
}
}
}
double MDBalancer::try_match(balance_state_t& state, mds_rank_t ex, double& maxex,
mds_rank_t im, double& maxim)
{
if (maxex <= 0 || maxim <= 0) return 0.0;
double howmuch = std::min(maxex, maxim);
dout(5) << " - mds." << ex << " exports " << howmuch << " to mds." << im << dendl;
if (ex == mds->get_nodeid())
state.targets[im] += howmuch;
state.exported[ex] += howmuch;
state.imported[im] += howmuch;
maxex -= howmuch;
maxim -= howmuch;
return howmuch;
}
void MDBalancer::queue_split(const CDir *dir, bool fast)
{
constexpr const auto &_func_ = __func__;
dout(10) << _func_ << " enqueuing " << *dir
<< " (fast=" << fast << ")" << dendl;
const dirfrag_t df = dir->dirfrag();
auto callback = [this, df](int r) {
if (split_pending.erase(df) == 0) {
// Someone beat me to it. This can happen in the fast splitting
// path, because we spawn two contexts, one with mds->timer and
// one with mds->queue_waiter. The loser can safely just drop
// out.
return;
}
if (mds->is_stopping()) {
// not a good time. This could have been (!mds->is_active())
// or at least (mds->is_stopping() || mds->is_stopped()), but
// is_stopped() is never true because an MDS respawns as soon as it's removed from the map;
// the narrow is_stopping check is to avoid potential regressions
// due to unknown coupling with other parts of the MDS (especially multiple ranks).
dout(5) << "ignoring the " << _func_ << " callback because the MDS state is '" << ceph_mds_state_name(mds->get_state()) << "'" << dendl;
return;
}
auto mdcache = mds->mdcache;
CDir *dir = mdcache->get_dirfrag(df);
if (!dir) {
dout(10) << "drop split on " << df << " because not in cache" << dendl;
return;
}
if (!dir->is_auth()) {
dout(10) << "drop split on " << df << " because non-auth" << dendl;
return;
}
// Pass on to MDCache: note that the split might still not
// happen if the checks in MDCache::can_fragment fail.
dout(10) << _func_ << " splitting " << *dir << dendl;
int bits = g_conf()->mds_bal_split_bits;
if (dir->inode->is_ephemeral_dist()) {
unsigned min_frag_bits = mdcache->get_ephemeral_dist_frag_bits();
if (df.frag.bits() + bits < min_frag_bits)
bits = min_frag_bits - df.frag.bits();
}
mdcache->split_dir(dir, bits);
};
auto ret = split_pending.insert(df);
bool is_new = ret.second;
if (fast) {
// Do the split ASAP: enqueue it in the MDSRank waiters which are
// run at the end of dispatching the current request
mds->queue_waiter(new MDSInternalContextWrapper(mds,
new LambdaContext(std::move(callback))));
} else if (is_new) {
// Set a timer to really do the split: we don't do it immediately
// so that bursts of ops on a directory have a chance to go through
// before we freeze it.
mds->timer.add_event_after(bal_fragment_interval,
new LambdaContext(std::move(callback)));
}
}
void MDBalancer::queue_merge(CDir *dir)
{
const auto frag = dir->dirfrag();
constexpr const auto &_func_ = __func__;
auto callback = [this, frag](int r) {
ceph_assert(frag.frag != frag_t());
// frag must be in this set because only one context is in flight
// for a given frag at a time (because merge_pending is checked before
// starting one), and this context is the only one that erases it.
merge_pending.erase(frag);
if (mds->is_stopping()) {
// not a good time. This could have been (!mds->is_active())
// or at least (mds->is_stopping() || mds->is_stopped()), but
// is_stopped() is never true because an MDS respawns as soon as it's removed from the map;
// the narrow is_stopping check is to avoid potential regressions
// due to unknown coupling with other parts of the MDS (especially multiple ranks).
dout(5) << "ignoring the " << _func_ << " callback because the MDS state is '" << ceph_mds_state_name(mds->get_state()) << "'" << dendl;
return;
}
auto mdcache = mds->mdcache;
CDir *dir = mdcache->get_dirfrag(frag);
if (!dir) {
dout(10) << "drop merge on " << frag << " because not in cache" << dendl;
return;
}
ceph_assert(dir->dirfrag() == frag);
if(!dir->is_auth()) {
dout(10) << "drop merge on " << *dir << " because lost auth" << dendl;
return;
}
dout(10) << "merging " << *dir << dendl;
CInode *diri = dir->get_inode();
unsigned min_frag_bits = 0;
if (diri->is_ephemeral_dist())
min_frag_bits = mdcache->get_ephemeral_dist_frag_bits();
frag_t fg = dir->get_frag();
while (fg.bits() > min_frag_bits) {
frag_t sibfg = fg.get_sibling();
auto&& [complete, sibs] = diri->get_dirfrags_under(sibfg);
if (!complete) {
dout(10) << " not all sibs under " << sibfg << " in cache (have " << sibs << ")" << dendl;
break;
}
bool all = true;
for (auto& sib : sibs) {
auto is_auth = sib->is_auth();
auto should_merge = sib->should_merge();
dout(20) << ": sib=" << *sib << ", is_auth=" << is_auth << ", should_merge="
<< should_merge << dendl;
if (!is_auth || !should_merge) {
all = false;
break;
}
}
if (!all) {
dout(10) << " not all sibs under " << sibfg << " " << sibs << " should_merge" << dendl;
break;
}
dout(10) << " all sibs under " << sibfg << " " << sibs << " should merge" << dendl;
fg = fg.parent();
}
if (fg != dir->get_frag())
mdcache->merge_dir(diri, fg);
};
if (merge_pending.count(frag) == 0) {
dout(20) << " enqueued dir " << *dir << dendl;
merge_pending.insert(frag);
mds->timer.add_event_after(bal_fragment_interval,
new LambdaContext(std::move(callback)));
} else {
dout(20) << " dir already in queue " << *dir << dendl;
}
}
void MDBalancer::prep_rebalance(int beat)
{
balance_state_t state;
if (g_conf()->mds_thrash_exports) {
//we're going to randomly export to all the mds in the cluster
set<mds_rank_t> up_mds;
mds->get_mds_map()->get_up_mds_set(up_mds);
for (const auto &rank : up_mds) {
state.targets[rank] = 0.0;
}
} else {
int cluster_size = mds->get_mds_map()->get_num_in_mds();
mds_rank_t whoami = mds->get_nodeid();
rebalance_time = clock::now();
dout(7) << "cluster loads are" << dendl;
mds->mdcache->migrator->clear_export_queue();
// rescale! turn my mds_load back into meta_load units
double load_fac = 1.0;
map<mds_rank_t, mds_load_t>::iterator m = mds_load.find(whoami);
if ((m != mds_load.end()) && (m->second.mds_load() > 0)) {
double metald = m->second.auth.meta_load();
double mdsld = m->second.mds_load();
load_fac = metald / mdsld;
dout(7) << " load_fac is " << load_fac
<< " <- " << m->second.auth << " " << metald
<< " / " << mdsld
<< dendl;
}
mds_meta_load.clear();
double total_load = 0.0;
multimap<double,mds_rank_t> load_map;
for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) {
mds_load_t& load = mds_load.at(i);
double l = load.mds_load() * load_fac;
mds_meta_load[i] = l;
if (whoami == 0)
dout(7) << " mds." << i
<< " " << load
<< " = " << load.mds_load()
<< " ~ " << l << dendl;
if (whoami == i) my_load = l;
total_load += l;
load_map.insert(pair<double,mds_rank_t>( l, i ));
}
// target load
target_load = total_load / (double)mds->mdsmap->get_num_mdss_in_rank_mask_bitset();
dout(7) << "my load " << my_load
<< " target " << target_load
<< " total " << total_load
<< dendl;
// under or over?
for (const auto& [load, rank] : load_map) {
if (test_rank_mask(rank) &&
load < target_load * (1.0 + g_conf()->mds_bal_min_rebalance)) {
dout(7) << " mds." << rank << " is underloaded or barely overloaded." << dendl;
mds_last_epoch_under_map[rank] = beat_epoch;
}
}
int last_epoch_under = mds_last_epoch_under_map[whoami];
if (last_epoch_under == beat_epoch) {
dout(7) << " i am underloaded or barely overloaded, doing nothing." << dendl;
return;
}
// am i over long enough?
if (last_epoch_under && beat_epoch - last_epoch_under < g_conf()->mds_bal_overload_epochs) {
dout(7) << " i am overloaded, but only for " << (beat_epoch - last_epoch_under) << " epochs" << dendl;
return;
}
dout(7) << " i am sufficiently overloaded" << dendl;
// first separate exporters and importers
multimap<double,mds_rank_t> importers;
multimap<double,mds_rank_t> exporters;
set<mds_rank_t> importer_set;
set<mds_rank_t> exporter_set;
for (multimap<double,mds_rank_t>::iterator it = load_map.begin();
it != load_map.end();
++it) {
if (it->first < target_load && test_rank_mask(it->second)) {
dout(15) << " mds." << it->second << " is importer" << dendl;
importers.insert(pair<double,mds_rank_t>(it->first,it->second));
importer_set.insert(it->second);
} else {
int mds_last_epoch_under = mds_last_epoch_under_map[it->second];
if (!(mds_last_epoch_under && beat_epoch - mds_last_epoch_under < 2)) {
dout(15) << " mds." << it->second << " is exporter" << dendl;
exporters.insert(pair<double,mds_rank_t>(it->first,it->second));
exporter_set.insert(it->second);
}
}
}
// determine load transfer mapping
if (true) {
// analyze import_map; do any matches i can
dout(15) << " matching exporters to import sources" << dendl;
// big -> small exporters
for (multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin();
ex != exporters.rend();
++ex) {
double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0;
double maxex = get_maxex(state, ex->second, ex_target_load);
if (maxex <= .001) continue;
// check importers. for now, just in arbitrary order (no intelligent matching).
for (map<mds_rank_t, float>::iterator im = mds_import_map[ex->second].begin();
im != mds_import_map[ex->second].end();
++im) {
double maxim = get_maxim(state, im->first, target_load);
if (maxim <= .001) continue;
try_match(state, ex->second, maxex, im->first, maxim);
if (maxex <= .001) break;
}
}
}
// old way
if (beat % 2 == 1) {
dout(15) << " matching big exporters to big importers" << dendl;
// big exporters to big importers
multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin();
multimap<double,mds_rank_t>::iterator im = importers.begin();
while (ex != exporters.rend() &&
im != importers.end()) {
double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0;
double maxex = get_maxex(state, ex->second, ex_target_load);
double maxim = get_maxim(state, im->second, target_load);
if (maxex < .001 || maxim < .001) break;
try_match(state, ex->second, maxex, im->second, maxim);
if (maxex <= .001) ++ex;
if (maxim <= .001) ++im;
}
} else { // new way
dout(15) << " matching small exporters to big importers" << dendl;
// small exporters to big importers
multimap<double,mds_rank_t>::iterator ex = exporters.begin();
multimap<double,mds_rank_t>::iterator im = importers.begin();
while (ex != exporters.end() &&
im != importers.end()) {
double ex_target_load = test_rank_mask(ex->second) ? target_load : 0.0;
double maxex = get_maxex(state, ex->second, ex_target_load);
double maxim = get_maxim(state, im->second, target_load);
if (maxex < .001 || maxim < .001) break;
try_match(state, ex->second, maxex, im->second, maxim);
if (maxex <= .001) ++ex;
if (maxim <= .001) ++im;
}
}
}
try_rebalance(state);
}
int MDBalancer::mantle_prep_rebalance()
{
balance_state_t state;
/* refresh balancer if it has changed */
if (bal_version != mds->mdsmap->get_balancer()) {
bal_version.assign("");
int r = localize_balancer();
if (r) return r;
/* only spam the cluster log from 1 mds on version changes */
if (mds->get_nodeid() == 0)
mds->clog->info() << "mantle balancer version changed: " << bal_version;
}
/* prepare for balancing */
int cluster_size = mds->get_mds_map()->get_num_in_mds();
rebalance_time = clock::now();
mds->mdcache->migrator->clear_export_queue();
/* fill in the metrics for each mds by grabbing load struct */
vector < map<string, double> > metrics (cluster_size);
for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) {
mds_load_t& load = mds_load.at(i);
metrics[i] = {{"auth.meta_load", load.auth.meta_load()},
{"all.meta_load", load.all.meta_load()},
{"req_rate", load.req_rate},
{"queue_len", load.queue_len},
{"cpu_load_avg", load.cpu_load_avg}};
}
/* execute the balancer */
Mantle mantle;
int ret = mantle.balance(bal_code, mds->get_nodeid(), metrics, state.targets);
dout(7) << " mantle decided that new targets=" << state.targets << dendl;
/* mantle doesn't know about cluster size, so check target len here */
if ((int) state.targets.size() != cluster_size)
return -CEPHFS_EINVAL;
else if (ret)
return ret;
try_rebalance(state);
return 0;
}
void MDBalancer::try_rebalance(balance_state_t& state)
{
if (g_conf()->mds_thrash_exports) {
dout(5) << "mds_thrash is on; not performing standard rebalance operation!"
<< dendl;
return;
}
// make a sorted list of my imports
multimap<double, CDir*> import_pop_map;
multimap<mds_rank_t, pair<CDir*, double> > import_from_map;
for (auto& dir : mds->mdcache->get_fullauth_subtrees()) {
CInode *diri = dir->get_inode();
if (diri->is_mdsdir())
continue;
if (diri->get_export_pin(false) != MDS_RANK_NONE)
continue;
if (dir->is_freezing() || dir->is_frozen())
continue; // export pbly already in progress
mds_rank_t from = diri->authority().first;
double pop = dir->pop_auth_subtree.meta_load();
if (g_conf()->mds_bal_idle_threshold > 0 &&
pop < g_conf()->mds_bal_idle_threshold &&
diri != mds->mdcache->get_root() &&
from != mds->get_nodeid()) {
dout(5) << " exporting idle (" << pop << ") import " << *dir
<< " back to mds." << from << dendl;
mds->mdcache->migrator->export_dir_nicely(dir, from);
continue;
}
dout(15) << " map: i imported " << *dir << " from " << from << dendl;
import_pop_map.insert(make_pair(pop, dir));
import_from_map.insert(make_pair(from, make_pair(dir, pop)));
}
// do my exports!
map<mds_rank_t, double> export_pop_map;
for (auto &it : state.targets) {
mds_rank_t target = it.first;
double amount = it.second;
if (amount < MIN_OFFLOAD) {
continue;
}
if (amount * 10 * state.targets.size() < target_load) {
continue;
}
dout(5) << "want to send " << amount << " to mds." << target
//<< " .. " << (*it).second << " * " << load_fac
<< " -> " << amount
<< dendl;//" .. fudge is " << fudge << dendl;
double& have = export_pop_map[target];