-
Notifications
You must be signed in to change notification settings - Fork 1.6k
/
threadpool_generic.cc
1771 lines (1450 loc) · 46.4 KB
/
threadpool_generic.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
/* Copyright (C) 2012, 2020, MariaDB Corporation.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA */
#if (defined HAVE_POOL_OF_THREADS) && !defined(EMBEDDED_LIBRARY)
#include "threadpool_generic.h"
#include "mariadb.h"
#include <violite.h>
#include <sql_priv.h>
#include <sql_class.h>
#include <my_pthread.h>
#include <scheduler.h>
#include <sql_connect.h>
#include <mysqld.h>
#include <debug_sync.h>
#include <time.h>
#include <sql_plist.h>
#include <threadpool.h>
#include <algorithm>
#ifdef HAVE_IOCP
#define OPTIONAL_IO_POLL_READ_PARAM this
#else
#define OPTIONAL_IO_POLL_READ_PARAM 0
#endif
static void io_poll_close(TP_file_handle fd)
{
#ifdef _WIN32
CloseHandle(fd);
#else
close(fd);
#endif
}
/** Maximum number of native events a listener can read in one go */
#define MAX_EVENTS 1024
/** Indicates that threadpool was initialized*/
static bool threadpool_started= false;
/*
Define PSI Keys for performance schema.
We have a mutex per group, worker threads, condition per worker thread,
and timer thread with its own mutex and condition.
*/
#ifdef HAVE_PSI_INTERFACE
static PSI_mutex_key key_group_mutex;
static PSI_mutex_key key_timer_mutex;
static PSI_mutex_info mutex_list[]=
{
{ &key_group_mutex, "group_mutex", 0},
{ &key_timer_mutex, "timer_mutex", PSI_FLAG_GLOBAL}
};
static PSI_cond_key key_worker_cond;
static PSI_cond_key key_timer_cond;
static PSI_cond_info cond_list[]=
{
{ &key_worker_cond, "worker_cond", 0},
{ &key_timer_cond, "timer_cond", PSI_FLAG_GLOBAL}
};
static PSI_thread_key key_worker_thread;
static PSI_thread_key key_timer_thread;
static PSI_thread_info thread_list[] =
{
{&key_worker_thread, "worker_thread", 0},
{&key_timer_thread, "timer_thread", PSI_FLAG_GLOBAL}
};
/* Macro to simplify performance schema registration */
#define PSI_register(X) \
if(PSI_server) PSI_server->register_ ## X("threadpool", X ## _list, array_elements(X ## _list))
#else
#define PSI_register(X) /* no-op */
#endif
thread_group_t *all_groups;
static uint group_count;
static Atomic_counter<uint32_t> shutdown_group_count;
/**
Used for printing "pool blocked" message, see
print_pool_blocked_message();
*/
static ulonglong pool_block_start;
/* Global timer for all groups */
struct pool_timer_t
{
mysql_mutex_t mutex;
mysql_cond_t cond;
volatile uint64 current_microtime;
std::atomic<uint64_t> next_timeout_check;
int tick_interval;
bool shutdown;
pthread_t timer_thread_id;
};
static pool_timer_t pool_timer;
static void queue_put(thread_group_t *thread_group, TP_connection_generic *connection);
static void queue_put(thread_group_t *thread_group, native_event *ev, int cnt);
static int wake_thread(thread_group_t *thread_group,bool due_to_stall);
static int wake_or_create_thread(thread_group_t *thread_group, bool due_to_stall=false);
static int create_worker(thread_group_t *thread_group, bool due_to_stall);
static void *worker_main(void *param);
static void check_stall(thread_group_t *thread_group);
static void set_next_timeout_check(ulonglong abstime);
static void print_pool_blocked_message(bool);
/**
Asynchronous network IO.
We use native edge-triggered network IO multiplexing facility.
This maps to different APIs on different Unixes.
Supported are currently Linux with epoll, Solaris with event ports,
OSX and BSD with kevent, Windows with IOCP. All those API's are used with one-shot flags
(the event is signalled once client has written something into the socket,
then socket is removed from the "poll-set" until the command is finished,
and we need to re-arm/re-register socket)
No implementation for poll/select is currently provided.
The API closely resembles all of the above mentioned platform APIs
and consists of following functions.
- io_poll_create()
Creates an io_poll descriptor
On Linux: epoll_create()
- io_poll_associate_fd(int poll_fd, TP_file_handle fd, void *data, void *opt)
Associate file descriptor with io poll descriptor
On Linux : epoll_ctl(..EPOLL_CTL_ADD))
- io_poll_disassociate_fd(TP_file_handle pollfd, TP_file_handle fd)
Associate file descriptor with io poll descriptor
On Linux: epoll_ctl(..EPOLL_CTL_DEL)
- io_poll_start_read(int poll_fd,int fd, void *data, void *opt)
The same as io_poll_associate_fd(), but cannot be used before
io_poll_associate_fd() was called.
On Linux : epoll_ctl(..EPOLL_CTL_MOD)
- io_poll_wait (TP_file_handle pollfd, native_event *native_events, int maxevents,
int timeout_ms)
wait until one or more descriptors added with io_poll_associate_fd()
or io_poll_start_read() becomes readable. Data associated with
descriptors can be retrieved from native_events array, using
native_event_get_userdata() function.
On Linux: epoll_wait()
*/
#if defined (__linux__)
#ifndef EPOLLRDHUP
/* Early 2.6 kernel did not have EPOLLRDHUP */
#define EPOLLRDHUP 0
#endif
static TP_file_handle io_poll_create()
{
return epoll_create(1);
}
int io_poll_associate_fd(TP_file_handle pollfd, TP_file_handle fd, void *data, void*)
{
struct epoll_event ev;
ev.data.u64= 0; /* Keep valgrind happy */
ev.data.ptr= data;
ev.events= EPOLLIN|EPOLLET|EPOLLERR|EPOLLRDHUP|EPOLLONESHOT;
return epoll_ctl(pollfd, EPOLL_CTL_ADD, fd, &ev);
}
int io_poll_start_read(TP_file_handle pollfd, TP_file_handle fd, void *data, void *)
{
struct epoll_event ev;
ev.data.u64= 0; /* Keep valgrind happy */
ev.data.ptr= data;
ev.events= EPOLLIN|EPOLLET|EPOLLERR|EPOLLRDHUP|EPOLLONESHOT;
return epoll_ctl(pollfd, EPOLL_CTL_MOD, fd, &ev);
}
int io_poll_disassociate_fd(TP_file_handle pollfd, TP_file_handle fd)
{
struct epoll_event ev;
return epoll_ctl(pollfd, EPOLL_CTL_DEL, fd, &ev);
}
/*
Wrapper around epoll_wait.
NOTE - in case of EINTR, it restarts with original timeout. Since we use
either infinite or 0 timeouts, this is not critical
*/
int io_poll_wait(TP_file_handle pollfd, native_event *native_events, int maxevents,
int timeout_ms)
{
int ret;
do
{
ret = epoll_wait(pollfd, native_events, maxevents, timeout_ms);
}
while(ret == -1 && errno == EINTR);
return ret;
}
static void *native_event_get_userdata(native_event *event)
{
return event->data.ptr;
}
#elif defined(HAVE_KQUEUE)
/*
NetBSD prior to 9.99.17 is incompatible with other BSDs, last parameter
in EV_SET macro (udata, user data) needs to be intptr_t, whereas it needs
to be void* everywhere else.
*/
#ifdef __NetBSD__
#include <sys/param.h>
# if !__NetBSD_Prereq__(9,99,17)
#define MY_EV_SET(a, b, c, d, e, f, g) EV_SET(a, b, c, d, e, f, (intptr_t)g)
# endif
#endif
#ifndef MY_EV_SET
#define MY_EV_SET(a, b, c, d, e, f, g) EV_SET(a, b, c, d, e, f, g)
#endif
TP_file_handle io_poll_create()
{
return kqueue();
}
int io_poll_start_read(TP_file_handle pollfd, TP_file_handle fd, void *data,void *)
{
struct kevent ke;
MY_EV_SET(&ke, fd, EVFILT_READ, EV_ADD|EV_ONESHOT,
0, 0, data);
return kevent(pollfd, &ke, 1, 0, 0, 0);
}
int io_poll_associate_fd(TP_file_handle pollfd, TP_file_handle fd, void *data,void *)
{
struct kevent ke;
MY_EV_SET(&ke, fd, EVFILT_READ, EV_ADD|EV_ONESHOT,
0, 0, data);
return io_poll_start_read(pollfd,fd, data, 0);
}
int io_poll_disassociate_fd(TP_file_handle pollfd, TP_file_handle fd)
{
struct kevent ke;
MY_EV_SET(&ke,fd, EVFILT_READ, EV_DELETE, 0, 0, 0);
return kevent(pollfd, &ke, 1, 0, 0, 0);
}
int io_poll_wait(TP_file_handle pollfd, struct kevent *events, int maxevents, int timeout_ms)
{
struct timespec ts;
int ret;
if (timeout_ms >= 0)
{
ts.tv_sec= timeout_ms/1000;
ts.tv_nsec= (timeout_ms%1000)*1000000;
}
do
{
ret= kevent(pollfd, 0, 0, events, maxevents,
(timeout_ms >= 0)?&ts:NULL);
}
while (ret == -1 && errno == EINTR);
return ret;
}
static void* native_event_get_userdata(native_event *event)
{
return (void *)event->udata;
}
#elif defined (__sun)
static TP_file_handle io_poll_create()
{
return port_create();
}
int io_poll_start_read(TP_file_handle pollfd, TP_file_handle fd, void *data, void *)
{
return port_associate(pollfd, PORT_SOURCE_FD, fd, POLLIN, data);
}
static int io_poll_associate_fd(TP_file_handle pollfd, TP_file_handle fd, void *data, void *)
{
return io_poll_start_read(pollfd, fd, data, 0);
}
int io_poll_disassociate_fd(TP_file_handle pollfd, TP_file_handle fd)
{
return port_dissociate(pollfd, PORT_SOURCE_FD, fd);
}
int io_poll_wait(TP_file_handle pollfd, native_event *events, int maxevents, int timeout_ms)
{
struct timespec ts;
int ret;
uint_t nget= 1;
if (timeout_ms >= 0)
{
ts.tv_sec= timeout_ms/1000;
ts.tv_nsec= (timeout_ms%1000)*1000000;
}
do
{
ret= port_getn(pollfd, events, maxevents, &nget,
(timeout_ms >= 0)?&ts:NULL);
}
while (ret == -1 && errno == EINTR);
DBUG_ASSERT(nget < INT_MAX);
return (int)nget;
}
static void* native_event_get_userdata(native_event *event)
{
return event->portev_user;
}
#elif defined(HAVE_IOCP)
static TP_file_handle io_poll_create()
{
return CreateIoCompletionPort(INVALID_HANDLE_VALUE, 0, 0, 0);
}
int io_poll_start_read(TP_file_handle pollfd, TP_file_handle fd, void *, void *opt)
{
static char c;
TP_connection_generic *con= (TP_connection_generic *)opt;
OVERLAPPED *overlapped= &con->overlapped;
if (con->vio_type == VIO_TYPE_NAMEDPIPE)
{
if (ReadFile(fd, &c, 0, NULL, overlapped))
return 0;
}
else
{
WSABUF buf;
buf.buf= &c;
buf.len= 0;
DWORD flags=0;
if (WSARecv((SOCKET)fd, &buf, 1,NULL, &flags,overlapped, NULL) == 0)
return 0;
}
if (GetLastError() == ERROR_IO_PENDING)
return 0;
return 1;
}
static int io_poll_associate_fd(TP_file_handle pollfd, TP_file_handle fd, void *data, void *opt)
{
HANDLE h= CreateIoCompletionPort(fd, pollfd, (ULONG_PTR)data, 0);
if (!h)
return -1;
return io_poll_start_read(pollfd,fd, 0, opt);
}
typedef LONG NTSTATUS;
typedef struct _IO_STATUS_BLOCK {
union {
NTSTATUS Status;
PVOID Pointer;
};
ULONG_PTR Information;
} IO_STATUS_BLOCK, * PIO_STATUS_BLOCK;
struct FILE_COMPLETION_INFORMATION {
HANDLE Port;
PVOID Key;
};
enum FILE_INFORMATION_CLASS {
FileReplaceCompletionInformation = 0x3D
};
typedef NTSTATUS(WINAPI* pNtSetInformationFile)(HANDLE, PIO_STATUS_BLOCK, PVOID, ULONG, FILE_INFORMATION_CLASS);
int io_poll_disassociate_fd(TP_file_handle pollfd, TP_file_handle fd)
{
static pNtSetInformationFile my_NtSetInformationFile = (pNtSetInformationFile)
GetProcAddress(GetModuleHandle("ntdll"), "NtSetInformationFile");
if (!my_NtSetInformationFile)
return -1; /* unexpected, we only support Windows 8.1+*/
IO_STATUS_BLOCK iosb{};
FILE_COMPLETION_INFORMATION fci{};
if (my_NtSetInformationFile(fd,&iosb,&fci,sizeof(fci),FileReplaceCompletionInformation))
return -1;
return 0;
}
int io_poll_wait(TP_file_handle pollfd, native_event *events, int maxevents, int timeout_ms)
{
ULONG n;
BOOL ok = GetQueuedCompletionStatusEx(pollfd, events,
maxevents, &n, timeout_ms, FALSE);
return ok ? (int)n : -1;
}
static void* native_event_get_userdata(native_event *event)
{
return (void *)event->lpCompletionKey;
}
#endif
/* Dequeue element from a workqueue */
static TP_connection_generic *queue_get(thread_group_t *thread_group)
{
DBUG_ENTER("queue_get");
thread_group->queue_event_count++;
TP_connection_generic *c;
for (int i=0; i < NQUEUES;i++)
{
c= thread_group->queues[i].pop_front();
if (c)
DBUG_RETURN(c);
}
DBUG_RETURN(0);
}
static TP_connection_generic* queue_get(thread_group_t* group, operation_origin origin)
{
auto ret = queue_get(group);
if (ret)
{
TP_INCREMENT_GROUP_COUNTER(group, dequeues[(int)origin]);
}
return ret;
}
static bool is_queue_empty(thread_group_t *thread_group)
{
for (int i=0; i < NQUEUES; i++)
{
if (!thread_group->queues[i].is_empty())
return false;
}
return true;
}
static void queue_init(thread_group_t *thread_group)
{
for (int i=0; i < NQUEUES; i++)
{
thread_group->queues[i].empty();
}
}
static void queue_put(thread_group_t *thread_group, native_event *ev, int cnt)
{
ulonglong now= threadpool_exact_stats?microsecond_interval_timer():pool_timer.current_microtime;
for(int i=0; i < cnt; i++)
{
TP_connection_generic *c = (TP_connection_generic *)native_event_get_userdata(&ev[i]);
c->enqueue_time= now;
thread_group->queues[c->priority].push_back(c);
}
}
/*
Handle wait timeout :
Find connections that have been idle for too long and kill them.
Also, recalculate time when next timeout check should run.
*/
static my_bool timeout_check(THD *thd, pool_timer_t *timer)
{
DBUG_ENTER("timeout_check");
if (thd->net.reading_or_writing == 1)
{
TP_connection_generic *connection= (TP_connection_generic *)thd->event_scheduler.data;
if (!connection || connection->state != TP_STATE_IDLE)
{
/*
Connection does not have scheduler data. This happens for example
if THD belongs to a different scheduler, that is listening to extra_port.
*/
DBUG_RETURN(0);
}
if(connection->abs_wait_timeout < timer->current_microtime)
{
tp_timeout_handler(connection);
}
else
{
if (connection->abs_wait_timeout < timer->current_microtime)
tp_timeout_handler(connection);
else
set_next_timeout_check(connection->abs_wait_timeout);
}
}
DBUG_RETURN(0);
}
/*
Timer thread.
Periodically, check if one of the thread groups is stalled. Stalls happen if
events are not being dequeued from the queue, or from the network, Primary
reason for stall can be a lengthy executing non-blocking request. It could
also happen that thread is waiting but wait_begin/wait_end is forgotten by
storage engine. Timer thread will create a new thread in group in case of
a stall.
Besides checking for stalls, timer thread is also responsible for terminating
clients that have been idle for longer than wait_timeout seconds.
TODO: Let the timer sleep for long time if there is no work to be done.
Currently it wakes up rather often on and idle server.
*/
static void* timer_thread(void *param)
{
uint i;
pool_timer_t* timer=(pool_timer_t *)param;
my_thread_init();
DBUG_ENTER("timer_thread");
timer->next_timeout_check.store(std::numeric_limits<uint64_t>::max(),
std::memory_order_relaxed);
timer->current_microtime= microsecond_interval_timer();
for(;;)
{
struct timespec ts;
int err;
set_timespec_nsec(ts,timer->tick_interval*1000000);
mysql_mutex_lock(&timer->mutex);
err= mysql_cond_timedwait(&timer->cond, &timer->mutex, &ts);
if (timer->shutdown)
{
mysql_mutex_unlock(&timer->mutex);
break;
}
if (err == ETIMEDOUT)
{
timer->current_microtime= microsecond_interval_timer();
/* Check stalls in thread groups */
for (i= 0; i < threadpool_max_size; i++)
{
if(all_groups[i].connection_count)
check_stall(&all_groups[i]);
}
/* Check if any client exceeded wait_timeout */
if (timer->next_timeout_check.load(std::memory_order_relaxed) <=
timer->current_microtime)
{
/* Reset next timeout check, it will be recalculated below */
timer->next_timeout_check.store(std::numeric_limits<uint64_t>::max(),
std::memory_order_relaxed);
server_threads.iterate(timeout_check, timer);
}
}
mysql_mutex_unlock(&timer->mutex);
}
mysql_mutex_destroy(&timer->mutex);
my_thread_end();
return NULL;
}
void check_stall(thread_group_t *thread_group)
{
mysql_mutex_lock(&thread_group->mutex);
/*
Bump priority for the low priority connections that spent too much
time in low prio queue.
*/
TP_connection_generic *c;
for (;;)
{
c= thread_group->queues[TP_PRIORITY_LOW].front();
if (c && pool_timer.current_microtime - c->enqueue_time > 1000ULL * threadpool_prio_kickup_timer)
{
thread_group->queues[TP_PRIORITY_LOW].remove(c);
thread_group->queues[TP_PRIORITY_HIGH].push_back(c);
}
else
break;
}
/*
Check if listener is present. If not, check whether any IO
events were dequeued since last time. If not, this means
listener is either in tight loop or thd_wait_begin()
was forgotten. Create a new worker(it will make itself listener).
*/
if (!thread_group->listener && !thread_group->io_event_count)
{
wake_or_create_thread(thread_group, true);
mysql_mutex_unlock(&thread_group->mutex);
return;
}
/* Reset io event count */
thread_group->io_event_count= 0;
/*
Check whether requests from the workqueue are being dequeued.
The stall detection and resolution works as follows:
1. There is a counter thread_group->queue_event_count for the number of
events removed from the queue. Timer resets the counter to 0 on each run.
2. Timer determines stall if this counter remains 0 since last check
and the queue is not empty.
3. Once timer determined a stall it sets thread_group->stalled flag and
wakes and idle worker (or creates a new one, subject to throttling).
4. The stalled flag is reset, when an event is dequeued.
Q : Will this handling lead to an unbound growth of threads, if queue
stalls permanently?
A : No. If queue stalls permanently, it is an indication for many very long
simultaneous queries. The maximum number of simultanoues queries is
max_connections, further we have threadpool_max_threads limit, upon which no
worker threads are created. So in case there is a flood of very long
queries, threadpool would slowly approach thread-per-connection behavior.
NOTE:
If long queries never wait, creation of the new threads is done by timer,
so it is slower than in real thread-per-connection. However if long queries
do wait and indicate that via thd_wait_begin/end callbacks, thread creation
will be faster.
*/
if (!is_queue_empty(thread_group) && !thread_group->queue_event_count)
{
thread_group->stalled= true;
TP_INCREMENT_GROUP_COUNTER(thread_group,stalls);
wake_or_create_thread(thread_group,true);
}
/* Reset queue event count */
thread_group->queue_event_count= 0;
mysql_mutex_unlock(&thread_group->mutex);
}
static void start_timer(pool_timer_t* timer)
{
DBUG_ENTER("start_timer");
mysql_mutex_init(key_timer_mutex,&timer->mutex, NULL);
mysql_cond_init(key_timer_cond, &timer->cond, NULL);
timer->shutdown = false;
mysql_thread_create(key_timer_thread, &timer->timer_thread_id, NULL,
timer_thread, timer);
DBUG_VOID_RETURN;
}
static void stop_timer(pool_timer_t *timer)
{
DBUG_ENTER("stop_timer");
mysql_mutex_lock(&timer->mutex);
timer->shutdown = true;
mysql_cond_signal(&timer->cond);
mysql_mutex_unlock(&timer->mutex);
pthread_join(timer->timer_thread_id, NULL);
DBUG_VOID_RETURN;
}
/**
Poll for socket events and distribute them to worker threads
In many case current thread will handle single event itself.
@return a ready connection, or NULL on shutdown
*/
static TP_connection_generic * listener(worker_thread_t *current_thread,
thread_group_t *thread_group)
{
DBUG_ENTER("listener");
TP_connection_generic *retval= NULL;
for(;;)
{
native_event ev[MAX_EVENTS];
int cnt;
if (thread_group->shutdown)
break;
cnt = io_poll_wait(thread_group->pollfd, ev, MAX_EVENTS, -1);
TP_INCREMENT_GROUP_COUNTER(thread_group, polls[(int)operation_origin::LISTENER]);
if (cnt <=0)
{
DBUG_ASSERT(thread_group->shutdown);
break;
}
mysql_mutex_lock(&thread_group->mutex);
if (thread_group->shutdown)
{
mysql_mutex_unlock(&thread_group->mutex);
break;
}
thread_group->io_event_count += cnt;
/*
We got some network events and need to make decisions : whether
listener hould handle events and whether or not any wake worker
threads so they can handle events.
Q1 : Should listener handle an event itself, or put all events into
queue and let workers handle the events?
Solution :
Generally, listener that handles events itself is preferable. We do not
want listener thread to change its state from waiting to running too
often, Since listener has just woken from poll, it better uses its time
slice and does some work. Besides, not handling events means they go to
the queue, and often to wake another worker must wake up to handle the
event. This is not good, as we want to avoid wakeups.
The downside of listener that also handles queries is that we can
potentially leave thread group for long time not picking the new
network events. It is not a major problem, because this stall will be
detected sooner or later by the timer thread. Still, relying on timer
is not always good, because it may "tick" too slow (large timer_interval)
We use following strategy to solve this problem - if queue was not empty
we suspect flood of network events and listener stays, Otherwise, it
handles a query.
Q2: If queue is not empty, how many workers to wake?
Solution:
We generally try to keep one thread per group active (threads handling
queries are considered active, unless they stuck in inside some "wait")
Thus, we will wake only one worker, and only if there is not active
threads currently,and listener is not going to handle a query. When we
don't wake, we hope that currently active threads will finish fast and
handle the queue. If this does not happen, timer thread will detect stall
and wake a worker.
NOTE: Currently nothing is done to detect or prevent long queuing times.
A solution for the future would be to give up "one active thread per
group" principle, if events stay in the queue for too long, and just wake
more workers.
*/
bool listener_picks_event=is_queue_empty(thread_group) && !threadpool_dedicated_listener;
queue_put(thread_group, ev, cnt);
if (listener_picks_event)
{
/* Handle the first event. */
retval= queue_get(thread_group, operation_origin::LISTENER);
mysql_mutex_unlock(&thread_group->mutex);
break;
}
if(thread_group->active_thread_count==0)
{
/* We added some work items to queue, now wake a worker. */
if(wake_thread(thread_group, false))
{
/*
Wake failed, hence groups has no idle threads. Now check if there are
any threads in the group except listener.
*/
if(thread_group->thread_count == 1)
{
/*
Currently there is no worker thread in the group, as indicated by
thread_count == 1 (this means listener is the only one thread in
the group).
The queue is not empty, and listener is not going to handle
events. In order to drain the queue, we create a worker here.
Alternatively, we could just rely on timer to detect stall, and
create thread, but waiting for timer would be an inefficient and
pointless delay.
*/
create_worker(thread_group, false);
}
}
}
mysql_mutex_unlock(&thread_group->mutex);
}
DBUG_RETURN(retval);
}
/**
Adjust thread counters in group or global
whenever thread is created or is about to exit
@param thread_group
@param count - 1, when new thread is created
-1, when thread is about to exit
*/
static void add_thread_count(thread_group_t *thread_group, int32 count)
{
thread_group->thread_count += count;
/* worker starts out and end in "active" state */
thread_group->active_thread_count += count;
tp_stats.num_worker_threads+= count;
}
/**
Creates a new worker thread.
thread_mutex must be held when calling this function
NOTE: in rare cases, the number of threads can exceed
threadpool_max_threads, because we need at least 2 threads
per group to prevent deadlocks (one listener + one worker)
*/
static int create_worker(thread_group_t *thread_group, bool due_to_stall)
{
pthread_t thread_id;
bool max_threads_reached= false;
int err;
DBUG_ENTER("create_worker");
if (tp_stats.num_worker_threads >= threadpool_max_threads
&& thread_group->thread_count >= 2)
{
err= 1;
max_threads_reached= true;
goto end;
}
err= mysql_thread_create(key_worker_thread, &thread_id,
thread_group->pthread_attr, worker_main, thread_group);
if (!err)
{
thread_group->last_thread_creation_time=microsecond_interval_timer();
statistic_increment(thread_created,&LOCK_status);
add_thread_count(thread_group, 1);
TP_INCREMENT_GROUP_COUNTER(thread_group,thread_creations);
if(due_to_stall)
{
TP_INCREMENT_GROUP_COUNTER(thread_group, thread_creations_due_to_stall);
}
}
else
{
my_errno= errno;
}
end:
if (err)
print_pool_blocked_message(max_threads_reached);
else
pool_block_start= 0; /* Reset pool blocked timer, if it was set */
DBUG_RETURN(err);
}
/**
Calculate microseconds throttling delay for thread creation.
The value depends on how many threads are already in the group:
small number of threads means no delay, the more threads the larger
the delay.
The actual values were not calculated using any scientific methods.
They just look right, and behave well in practice.
*/
#define THROTTLING_FACTOR (threadpool_stall_limit/std::max(DEFAULT_THREADPOOL_STALL_LIMIT,threadpool_stall_limit))
static ulonglong microsecond_throttling_interval(thread_group_t *thread_group)
{
int count= thread_group->thread_count;
if (count < 1+ (int)threadpool_oversubscribe)
return 0;
if (count < 8)
return 50*1000*THROTTLING_FACTOR;
if(count < 16)
return 100*1000*THROTTLING_FACTOR;
return 200*100*THROTTLING_FACTOR;
}
/**
Wakes a worker thread, or creates a new one.
Worker creation is throttled, so we avoid too many threads
to be created during the short time.
*/
static int wake_or_create_thread(thread_group_t *thread_group, bool due_to_stall)
{
DBUG_ENTER("wake_or_create_thread");
if (thread_group->shutdown)
DBUG_RETURN(0);
if (wake_thread(thread_group, due_to_stall) == 0)
{
DBUG_RETURN(0);
}
if (thread_group->thread_count > thread_group->connection_count)
DBUG_RETURN(-1);
if (thread_group->active_thread_count == 0)
{
/*
We're better off creating a new thread here with no delay, either there
are no workers at all, or they all are all blocking and there was no
idle thread to wakeup. Smells like a potential deadlock or very slowly
executing requests, e.g sleeps or user locks.
*/
DBUG_RETURN(create_worker(thread_group, due_to_stall));
}
ulonglong now = microsecond_interval_timer();
ulonglong time_since_last_thread_created =
(now - thread_group->last_thread_creation_time);
/* Throttle thread creation. */
if (time_since_last_thread_created >
microsecond_throttling_interval(thread_group))
{
DBUG_RETURN(create_worker(thread_group, due_to_stall));
}
TP_INCREMENT_GROUP_COUNTER(thread_group,throttles);
DBUG_RETURN(-1);
}
int thread_group_init(thread_group_t *thread_group, pthread_attr_t* thread_attr)
{
DBUG_ENTER("thread_group_init");
thread_group->pthread_attr = thread_attr;
mysql_mutex_init(key_group_mutex, &thread_group->mutex, NULL);
thread_group->pollfd= INVALID_HANDLE_VALUE;
thread_group->shutdown_pipe[0]= -1;
thread_group->shutdown_pipe[1]= -1;
queue_init(thread_group);