1
/*****************************************************************************
3
Copyright (c) 1995, 2010, Innobase Oy. All Rights Reserved.
4
Copyright (c) 2008, 2009 Google Inc.
5
Copyright (c) 2009, Percona Inc.
7
Portions of this file contain modifications contributed and copyrighted by
8
Google, Inc. Those modifications are gratefully acknowledged and are described
9
briefly in the InnoDB documentation. The contributions by Google are
10
incorporated with their permission, and subject to the conditions contained in
11
the file COPYING.Google.
13
Portions of this file contain modifications contributed and copyrighted
14
by Percona Inc.. Those modifications are
15
gratefully acknowledged and are described briefly in the InnoDB
16
documentation. The contributions by Percona Inc. are incorporated with
17
their permission, and subject to the conditions contained in the file
20
This program is free software; you can redistribute it and/or modify it under
21
the terms of the GNU General Public License as published by the Free Software
22
Foundation; version 2 of the License.
24
This program is distributed in the hope that it will be useful, but WITHOUT
25
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
26
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
28
You should have received a copy of the GNU General Public License along with
29
this program; if not, write to the Free Software Foundation, Inc., 51 Franklin
30
St, Fifth Floor, Boston, MA 02110-1301 USA
32
*****************************************************************************/
34
/**************************************************//**
36
The database server main program
38
NOTE: SQL Server 7 uses something which the documentation
39
calls user mode scheduled threads (UMS threads). One such
40
thread is usually allocated per processor. Win32
41
documentation does not know any UMS threads, which suggests
42
that the concept is internal to SQL Server 7. It may mean that
43
SQL Server 7 does all the scheduling of threads itself, even
44
in i/o waits. We should maybe modify InnoDB to use the same
45
technique, because thread switches within NT may be too slow.
47
SQL Server 7 also mentions fibers, which are cooperatively
48
scheduled threads. They can boost performance by 5 %,
49
according to the Delaney and Soukup's book.
51
Windows 2000 will have something called thread pooling
52
(see msdn website), which we could possibly use.
54
Another possibility could be to use some very fast user space
55
thread library. This might confuse NT though.
57
Created 10/8/1995 Heikki Tuuri
58
*******************************************************/
68
#include "sync0sync.h"
72
#include "pars0pars.h"
74
#include "lock0lock.h"
75
#include "trx0purge.h"
76
#include "ibuf0ibuf.h"
80
#include "dict0load.h"
81
#include "dict0boot.h"
82
#include "srv0start.h"
83
#include "row0mysql.h"
84
#include "ha_prototypes.h"
86
#include "os0sync.h" /* for HAVE_ATOMIC_BUILTINS */
88
/* This is set to TRUE if the MySQL user has set it in MySQL; currently
89
affects only FOREIGN KEY definition parsing */
90
UNIV_INTERN ibool srv_lower_case_table_names = FALSE;
92
/* The following counter is incremented whenever there is some user activity
94
UNIV_INTERN ulint srv_activity_count = 0;
96
/* The following is the maximum allowed duration of a lock wait. */
97
UNIV_INTERN ulint srv_fatal_semaphore_wait_threshold = 600;
99
/* How much data manipulation language (DML) statements need to be delayed,
100
in microseconds, in order to reduce the lagging of the purge thread. */
101
UNIV_INTERN ulint srv_dml_needed_delay = 0;
103
UNIV_INTERN ibool srv_lock_timeout_active = FALSE;
104
UNIV_INTERN ibool srv_monitor_active = FALSE;
105
UNIV_INTERN ibool srv_error_monitor_active = FALSE;
107
UNIV_INTERN const char* srv_main_thread_op_info = "";
109
/** Prefix used by MySQL to indicate pre-5.1 table name encoding */
110
UNIV_INTERN const char srv_mysql50_table_name_prefix[9] = "#mysql50#";
112
/* Server parameters which are read from the initfile */
114
/* The following three are dir paths which are catenated before file
115
names, where the file name itself may also contain a path */
117
UNIV_INTERN char* srv_data_home = NULL;
118
#ifdef UNIV_LOG_ARCHIVE
119
UNIV_INTERN char* srv_arch_dir = NULL;
120
#endif /* UNIV_LOG_ARCHIVE */
122
/** store to its own file each table created by an user; data
123
dictionary tables are in the system tablespace 0 */
124
UNIV_INTERN my_bool srv_file_per_table;
125
/** The file format to use on new *.ibd files. */
126
UNIV_INTERN ulint srv_file_format = 0;
127
/** Whether to check file format during startup. A value of
128
DICT_TF_FORMAT_MAX + 1 means no checking ie. FALSE. The default is to
129
set it to the highest format we support. */
130
UNIV_INTERN ulint srv_max_file_format_at_startup = DICT_TF_FORMAT_MAX;
132
#if DICT_TF_FORMAT_51
133
# error "DICT_TF_FORMAT_51 must be 0!"
135
/** Place locks to records only i.e. do not use next-key locking except
136
on duplicate key checking and foreign key checking */
137
UNIV_INTERN ibool srv_locks_unsafe_for_binlog = FALSE;
139
/* If this flag is TRUE, then we will use the native aio of the
140
OS (provided we compiled Innobase with it in), otherwise we will
141
use simulated aio we build below with threads.
142
Currently we support native aio on windows and linux */
143
UNIV_INTERN my_bool srv_use_native_aio = TRUE;
146
/* Windows native condition variables. We use runtime loading / function
147
pointers, because they are not available on Windows Server 2003 and
150
We use condition for events on Windows if possible, even if os_event
151
resembles Windows kernel event object well API-wise. The reason is
152
performance, kernel objects are heavyweights and WaitForSingleObject() is a
153
performance killer causing calling thread to context switch. Besides, Innodb
154
is preallocating large number (often millions) of os_events. With kernel event
155
objects it takes a big chunk out of non-paged pool, which is better suited
156
for tasks like IO than for storing idle event objects. */
157
UNIV_INTERN ibool srv_use_native_conditions = FALSE;
160
UNIV_INTERN ulint srv_n_data_files = 0;
161
UNIV_INTERN char** srv_data_file_names = NULL;
162
/* size in database pages */
163
UNIV_INTERN ulint* srv_data_file_sizes = NULL;
165
/* if TRUE, then we auto-extend the last data file */
166
UNIV_INTERN ibool srv_auto_extend_last_data_file = FALSE;
167
/* if != 0, this tells the max size auto-extending may increase the
168
last data file size */
169
UNIV_INTERN ulint srv_last_file_size_max = 0;
170
/* If the last data file is auto-extended, we add this
171
many pages to it at a time */
172
UNIV_INTERN unsigned int srv_auto_extend_increment = 8;
173
UNIV_INTERN ulint* srv_data_file_is_raw_partition = NULL;
175
/* If the following is TRUE we do not allow inserts etc. This protects
176
the user from forgetting the 'newraw' keyword to my.cnf */
178
UNIV_INTERN ibool srv_created_new_raw = FALSE;
180
UNIV_INTERN char** srv_log_group_home_dirs = NULL;
182
UNIV_INTERN ulint srv_n_log_groups = ULINT_MAX;
183
UNIV_INTERN ulint srv_n_log_files = ULINT_MAX;
184
/* size in database pages */
185
UNIV_INTERN ulint srv_log_file_size = ULINT_MAX;
186
/* size in database pages */
187
UNIV_INTERN ulint srv_log_buffer_size = ULINT_MAX;
188
UNIV_INTERN ulong srv_flush_log_at_trx_commit = 1;
190
/* Try to flush dirty pages so as to avoid IO bursts at
192
UNIV_INTERN bool srv_adaptive_flushing = TRUE;
194
/** Maximum number of times allowed to conditionally acquire
195
mutex before switching to blocking wait on the mutex */
196
#define MAX_MUTEX_NOWAIT 20
198
/** Check whether the number of failed nonblocking mutex
199
acquisition attempts exceeds maximum allowed value. If so,
200
srv_printf_innodb_monitor() will request mutex acquisition
201
with mutex_enter(), which will wait until it gets the mutex. */
202
#define MUTEX_NOWAIT(mutex_skipped) ((mutex_skipped) < MAX_MUTEX_NOWAIT)
204
/** The sort order table of the MySQL latin1_swedish_ci character set
206
#if defined(BUILD_DRIZZLE)
207
UNIV_INTERN const byte srv_latin1_ordering[256] /* The sort order table of the latin1
208
character set. The following table is
209
the MySQL order as of Feb 10th, 2002 */
211
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07
212
, 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F
213
, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17
214
, 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F
215
, 0x20, 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27
216
, 0x28, 0x29, 0x2A, 0x2B, 0x2C, 0x2D, 0x2E, 0x2F
217
, 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37
218
, 0x38, 0x39, 0x3A, 0x3B, 0x3C, 0x3D, 0x3E, 0x3F
219
, 0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47
220
, 0x48, 0x49, 0x4A, 0x4B, 0x4C, 0x4D, 0x4E, 0x4F
221
, 0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57
222
, 0x58, 0x59, 0x5A, 0x5B, 0x5C, 0x5D, 0x5E, 0x5F
223
, 0x60, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47
224
, 0x48, 0x49, 0x4A, 0x4B, 0x4C, 0x4D, 0x4E, 0x4F
225
, 0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57
226
, 0x58, 0x59, 0x5A, 0x7B, 0x7C, 0x7D, 0x7E, 0x7F
227
, 0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87
228
, 0x88, 0x89, 0x8A, 0x8B, 0x8C, 0x8D, 0x8E, 0x8F
229
, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97
230
, 0x98, 0x99, 0x9A, 0x9B, 0x9C, 0x9D, 0x9E, 0x9F
231
, 0xA0, 0xA1, 0xA2, 0xA3, 0xA4, 0xA5, 0xA6, 0xA7
232
, 0xA8, 0xA9, 0xAA, 0xAB, 0xAC, 0xAD, 0xAE, 0xAF
233
, 0xB0, 0xB1, 0xB2, 0xB3, 0xB4, 0xB5, 0xB6, 0xB7
234
, 0xB8, 0xB9, 0xBA, 0xBB, 0xBC, 0xBD, 0xBE, 0xBF
235
, 0x41, 0x41, 0x41, 0x41, 0x5C, 0x5B, 0x5C, 0x43
236
, 0x45, 0x45, 0x45, 0x45, 0x49, 0x49, 0x49, 0x49
237
, 0x44, 0x4E, 0x4F, 0x4F, 0x4F, 0x4F, 0x5D, 0xD7
238
, 0xD8, 0x55, 0x55, 0x55, 0x59, 0x59, 0xDE, 0xDF
239
, 0x41, 0x41, 0x41, 0x41, 0x5C, 0x5B, 0x5C, 0x43
240
, 0x45, 0x45, 0x45, 0x45, 0x49, 0x49, 0x49, 0x49
241
, 0x44, 0x4E, 0x4F, 0x4F, 0x4F, 0x4F, 0x5D, 0xF7
242
, 0xD8, 0x55, 0x55, 0x55, 0x59, 0x59, 0xDE, 0xFF
245
UNIV_INTERN const byte* srv_latin1_ordering;
246
#endif /* BUILD_DRIZZLE */
249
/* use os/external memory allocator */
250
UNIV_INTERN my_bool srv_use_sys_malloc = TRUE;
251
/* requested size in kilobytes */
252
UNIV_INTERN ulint srv_buf_pool_size = ULINT_MAX;
253
/* requested number of buffer pool instances */
254
UNIV_INTERN ulint srv_buf_pool_instances = 1;
255
/* previously requested size */
256
UNIV_INTERN ulint srv_buf_pool_old_size;
257
/* current size in kilobytes */
258
UNIV_INTERN ulint srv_buf_pool_curr_size = 0;
260
UNIV_INTERN ulint srv_mem_pool_size = ULINT_MAX;
261
UNIV_INTERN ulint srv_lock_table_size = ULINT_MAX;
263
/* This parameter is deprecated. Use srv_n_io_[read|write]_threads
265
UNIV_INTERN ulint srv_n_file_io_threads = ULINT_MAX;
266
UNIV_INTERN ulint srv_n_read_io_threads = ULINT_MAX;
267
UNIV_INTERN ulint srv_n_write_io_threads = ULINT_MAX;
269
/* User settable value of the number of pages that must be present
270
in the buffer cache and accessed sequentially for InnoDB to trigger a
271
readahead request. */
272
UNIV_INTERN ulong srv_read_ahead_threshold = 56;
274
#ifdef UNIV_LOG_ARCHIVE
275
UNIV_INTERN ibool srv_log_archive_on = FALSE;
276
UNIV_INTERN ibool srv_archive_recovery = 0;
277
UNIV_INTERN ib_uint64_t srv_archive_recovery_limit_lsn;
278
#endif /* UNIV_LOG_ARCHIVE */
280
/* This parameter is used to throttle the number of insert buffers that are
281
merged in a batch. By increasing this parameter on a faster disk you can
282
possibly reduce the number of I/O operations performed to complete the
283
merge operation. The value of this parameter is used as is by the
284
background loop when the system is idle (low load), on a busy system
285
the parameter is scaled down by a factor of 4, this is to avoid putting
286
a heavier load on the I/O sub system. */
288
UNIV_INTERN ulong srv_insert_buffer_batch_size = 20;
290
UNIV_INTERN char* srv_file_flush_method_str = NULL;
291
UNIV_INTERN ulint srv_unix_file_flush_method = SRV_UNIX_FSYNC;
292
UNIV_INTERN ulint srv_win_file_flush_method = SRV_WIN_IO_UNBUFFERED;
294
UNIV_INTERN ulint srv_max_n_open_files = 300;
296
/* Number of IO operations per second the server can do */
297
UNIV_INTERN ulong srv_io_capacity = 200;
299
/* The InnoDB main thread tries to keep the ratio of modified pages
300
in the buffer pool to all database pages in the buffer pool smaller than
301
the following number. But it is not guaranteed that the value stays below
302
that during a time of heavy update/insert activity. */
304
UNIV_INTERN ulong srv_max_buf_pool_modified_pct = 75;
306
/* the number of purge threads to use from the worker pool (currently 0 or 1).*/
307
UNIV_INTERN ulong srv_n_purge_threads = 0;
309
/* the number of records to purge in one batch */
310
UNIV_INTERN ulong srv_purge_batch_size = 20;
312
/* variable counts amount of data read in total (in bytes) */
313
UNIV_INTERN ulint srv_data_read = 0;
315
/* here we count the amount of data written in total (in bytes) */
316
UNIV_INTERN ulint srv_data_written = 0;
318
/* the number of the log write requests done */
319
UNIV_INTERN ulint srv_log_write_requests = 0;
321
/* the number of physical writes to the log performed */
322
UNIV_INTERN ulint srv_log_writes = 0;
324
/* amount of data written to the log files in bytes */
325
UNIV_INTERN ulint srv_os_log_written = 0;
327
/* amount of writes being done to the log files */
328
UNIV_INTERN ulint srv_os_log_pending_writes = 0;
330
/* we increase this counter, when there we don't have enough space in the
331
log buffer and have to flush it */
332
UNIV_INTERN ulint srv_log_waits = 0;
334
/* this variable counts the amount of times, when the doublewrite buffer
336
UNIV_INTERN ulint srv_dblwr_writes = 0;
338
/* here we store the number of pages that have been flushed to the
339
doublewrite buffer */
340
UNIV_INTERN ulint srv_dblwr_pages_written = 0;
342
/* in this variable we store the number of write requests issued */
343
UNIV_INTERN ulint srv_buf_pool_write_requests = 0;
345
/* here we store the number of times when we had to wait for a free page
346
in the buffer pool. It happens when the buffer pool is full and we need
347
to make a flush, in order to be able to read or create a page. */
348
UNIV_INTERN ulint srv_buf_pool_wait_free = 0;
350
/* variable to count the number of pages that were written from buffer
352
UNIV_INTERN ulint srv_buf_pool_flushed = 0;
354
/** Number of buffer pool reads that led to the
355
reading of a disk page */
356
UNIV_INTERN ulint srv_buf_pool_reads = 0;
358
/* structure to pass status variables to MySQL */
359
UNIV_INTERN export_struc export_vars;
361
/* If the following is != 0 we do not allow inserts etc. This protects
362
the user from forgetting the innodb_force_recovery keyword to my.cnf */
364
UNIV_INTERN ulint srv_force_recovery = 0;
365
/*-----------------------*/
366
/* We are prepared for a situation that we have this many threads waiting for
367
a semaphore inside InnoDB. innobase_start_or_create_for_mysql() sets the
370
UNIV_INTERN ulint srv_max_n_threads = 0;
372
/* The following controls how many threads we let inside InnoDB concurrently:
373
threads waiting for locks are not counted into the number because otherwise
374
we could get a deadlock. MySQL creates a thread for each user session, and
375
semaphore contention and convoy problems can occur withput this restriction.
376
Value 10 should be good if there are less than 4 processors + 4 disks in the
377
computer. Bigger computers need bigger values. Value 0 will disable the
378
concurrency check. */
380
UNIV_INTERN ulong srv_thread_concurrency = 0;
382
/* this mutex protects srv_conc data structures */
383
UNIV_INTERN os_fast_mutex_t srv_conc_mutex;
384
/* number of transactions that have declared_to_be_inside_innodb set.
385
It used to be a non-error for this value to drop below zero temporarily.
386
This is no longer true. We'll, however, keep the lint datatype to add
387
assertions to catch any corner cases that we may have missed. */
388
UNIV_INTERN lint srv_conc_n_threads = 0;
389
/* number of OS threads waiting in the FIFO for a permission to enter
391
UNIV_INTERN ulint srv_conc_n_waiting_threads = 0;
393
typedef struct srv_conc_slot_struct srv_conc_slot_t;
394
struct srv_conc_slot_struct{
395
os_event_t event; /*!< event to wait */
396
ibool reserved; /*!< TRUE if slot
398
ibool wait_ended; /*!< TRUE when another
399
thread has already set
401
thread in this slot is
403
reserved may still be
404
TRUE at that point */
405
UT_LIST_NODE_T(srv_conc_slot_t) srv_conc_queue; /*!< queue node */
408
/* queue of threads waiting to get in */
409
UNIV_INTERN UT_LIST_BASE_NODE_T(srv_conc_slot_t) srv_conc_queue;
410
/* array of wait slots */
411
UNIV_INTERN srv_conc_slot_t* srv_conc_slots;
413
/* Number of times a thread is allowed to enter InnoDB within the same
414
SQL query after it has once got the ticket at srv_conc_enter_innodb */
415
#define SRV_FREE_TICKETS_TO_ENTER srv_n_free_tickets_to_enter
416
#define SRV_THREAD_SLEEP_DELAY srv_thread_sleep_delay
417
/*-----------------------*/
418
/* If the following is set to 1 then we do not run purge and insert buffer
419
merge to completion before shutdown. If it is set to 2, do not even flush the
420
buffer pool to data files at the shutdown: we effectively 'crash'
421
InnoDB (but lose no committed transactions). */
422
UNIV_INTERN ulint srv_fast_shutdown = 0;
424
/* Generate a innodb_status.<pid> file */
425
UNIV_INTERN ibool srv_innodb_status = FALSE;
427
/* When estimating number of different key values in an index, sample
428
this many index pages */
429
UNIV_INTERN ib_uint64_t srv_stats_sample_pages = 8;
431
UNIV_INTERN ibool srv_use_doublewrite_buf = TRUE;
432
UNIV_INTERN ibool srv_use_checksums = TRUE;
434
UNIV_INTERN ulong srv_replication_delay = 0;
436
/*-------------------------------------------*/
437
UNIV_INTERN ulong srv_n_spin_wait_rounds = 30;
438
UNIV_INTERN ulong srv_n_free_tickets_to_enter = 500;
439
UNIV_INTERN ulong srv_thread_sleep_delay = 10000;
440
UNIV_INTERN ulong srv_spin_wait_delay = 6;
441
UNIV_INTERN ibool srv_priority_boost = TRUE;
444
UNIV_INTERN ibool srv_print_thread_releases = FALSE;
445
UNIV_INTERN ibool srv_print_lock_waits = FALSE;
446
UNIV_INTERN ibool srv_print_buf_io = FALSE;
447
UNIV_INTERN ibool srv_print_log_io = FALSE;
448
UNIV_INTERN ibool srv_print_latch_waits = FALSE;
449
#endif /* UNIV_DEBUG */
451
UNIV_INTERN ulint srv_n_rows_inserted = 0;
452
UNIV_INTERN ulint srv_n_rows_updated = 0;
453
UNIV_INTERN ulint srv_n_rows_deleted = 0;
454
UNIV_INTERN ulint srv_n_rows_read = 0;
456
static ulint srv_n_rows_inserted_old = 0;
457
static ulint srv_n_rows_updated_old = 0;
458
static ulint srv_n_rows_deleted_old = 0;
459
static ulint srv_n_rows_read_old = 0;
461
UNIV_INTERN ulint srv_n_lock_wait_count = 0;
462
UNIV_INTERN ulint srv_n_lock_wait_current_count = 0;
463
UNIV_INTERN ib_int64_t srv_n_lock_wait_time = 0;
464
UNIV_INTERN ulint srv_n_lock_max_wait_time = 0;
466
UNIV_INTERN ulint srv_truncated_status_writes = 0;
469
Set the following to 0 if you want InnoDB to write messages on
470
stderr on startup/shutdown
472
UNIV_INTERN ibool srv_print_verbose_log = TRUE;
473
UNIV_INTERN ibool srv_print_innodb_monitor = FALSE;
474
UNIV_INTERN ibool srv_print_innodb_lock_monitor = FALSE;
475
UNIV_INTERN ibool srv_print_innodb_tablespace_monitor = FALSE;
476
UNIV_INTERN ibool srv_print_innodb_table_monitor = FALSE;
478
/* Array of English strings describing the current state of an
479
i/o handler thread */
481
UNIV_INTERN const char* srv_io_thread_op_info[SRV_MAX_N_IO_THREADS];
482
UNIV_INTERN const char* srv_io_thread_function[SRV_MAX_N_IO_THREADS];
484
UNIV_INTERN time_t srv_last_monitor_time;
486
UNIV_INTERN mutex_t srv_innodb_monitor_mutex;
488
/* Mutex for locking srv_monitor_file */
489
UNIV_INTERN mutex_t srv_monitor_file_mutex;
491
#ifdef UNIV_PFS_MUTEX
492
/* Key to register kernel_mutex with performance schema */
493
UNIV_INTERN mysql_pfs_key_t kernel_mutex_key;
494
/* Key to register srv_innodb_monitor_mutex with performance schema */
495
UNIV_INTERN mysql_pfs_key_t srv_innodb_monitor_mutex_key;
496
/* Key to register srv_monitor_file_mutex with performance schema */
497
UNIV_INTERN mysql_pfs_key_t srv_monitor_file_mutex_key;
498
/* Key to register srv_dict_tmpfile_mutex with performance schema */
499
UNIV_INTERN mysql_pfs_key_t srv_dict_tmpfile_mutex_key;
500
/* Key to register the mutex with performance schema */
501
UNIV_INTERN mysql_pfs_key_t srv_misc_tmpfile_mutex_key;
502
#endif /* UNIV_PFS_MUTEX */
504
/* Temporary file for innodb monitor output */
505
UNIV_INTERN FILE* srv_monitor_file;
506
/* Mutex for locking srv_dict_tmpfile.
507
This mutex has a very high rank; threads reserving it should not
508
be holding any InnoDB latches. */
509
UNIV_INTERN mutex_t srv_dict_tmpfile_mutex;
510
/* Temporary file for output from the data dictionary */
511
UNIV_INTERN FILE* srv_dict_tmpfile;
512
/* Mutex for locking srv_misc_tmpfile.
513
This mutex has a very low rank; threads reserving it should not
514
acquire any further latches or sleep before releasing this one. */
515
UNIV_INTERN mutex_t srv_misc_tmpfile_mutex;
516
/* Temporary file for miscellanous diagnostic output */
517
UNIV_INTERN FILE* srv_misc_tmpfile;
519
UNIV_INTERN ulint srv_main_thread_process_no = 0;
520
UNIV_INTERN ulint srv_main_thread_id = 0;
522
/* The following count work done by srv_master_thread. */
524
/* Iterations by the 'once per second' loop. */
525
static ulint srv_main_1_second_loops = 0;
526
/* Calls to sleep by the 'once per second' loop. */
527
static ulint srv_main_sleeps = 0;
528
/* Iterations by the 'once per 10 seconds' loop. */
529
static ulint srv_main_10_second_loops = 0;
530
/* Iterations of the loop bounded by the 'background_loop' label. */
531
static ulint srv_main_background_loops = 0;
532
/* Iterations of the loop bounded by the 'flush_loop' label. */
533
static ulint srv_main_flush_loops = 0;
534
/* Log writes involving flush. */
535
static ulint srv_log_writes_and_flush = 0;
537
/* This is only ever touched by the master thread. It records the
538
time when the last flush of log file has happened. The master
539
thread ensures that we flush the log files at least once per
541
static time_t srv_last_log_flush_time;
543
/* The master thread performs various tasks based on the current
544
state of IO activity and the level of IO utilization is past
545
intervals. Following macros define thresholds for these conditions. */
546
#define SRV_PEND_IO_THRESHOLD (PCT_IO(3))
547
#define SRV_RECENT_IO_ACTIVITY (PCT_IO(5))
548
#define SRV_PAST_IO_ACTIVITY (PCT_IO(200))
551
IMPLEMENTATION OF THE SERVER MAIN PROGRAM
552
=========================================
554
There is the following analogue between this database
555
server and an operating system kernel:
557
DB concept equivalent OS concept
558
---------- ---------------------
559
transaction -- process;
561
query thread -- thread;
566
the rollback state -- kill signal delivered to a process;
570
query thread execution:
571
(a) without kernel mutex
572
reserved -- process executing in user mode;
573
(b) with kernel mutex reserved
574
-- process executing in kernel mode;
576
The server is controlled by a master thread which runs at
577
a priority higher than normal, that is, higher than user threads.
578
It sleeps most of the time, and wakes up, say, every 300 milliseconds,
579
to check whether there is anything happening in the server which
580
requires intervention of the master thread. Such situations may be,
581
for example, when flushing of dirty blocks is needed in the buffer
582
pool or old version of database rows have to be cleaned away.
584
The threads which we call user threads serve the queries of
585
the clients and input from the console of the server.
586
They run at normal priority. The server may have several
587
communications endpoints. A dedicated set of user threads waits
588
at each of these endpoints ready to receive a client request.
589
Each request is taken by a single user thread, which then starts
590
processing and, when the result is ready, sends it to the client
591
and returns to wait at the same endpoint the thread started from.
593
So, we do not have dedicated communication threads listening at
594
the endpoints and dealing the jobs to dedicated worker threads.
595
Our architecture saves one thread swithch per request, compared
596
to the solution with dedicated communication threads
597
which amounts to 15 microseconds on 100 MHz Pentium
598
running NT. If the client
599
is communicating over a network, this saving is negligible, but
600
if the client resides in the same machine, maybe in an SMP machine
601
on a different processor from the server thread, the saving
602
can be important as the threads can communicate over shared
603
memory with an overhead of a few microseconds.
605
We may later implement a dedicated communication thread solution
606
for those endpoints which communicate over a network.
608
Our solution with user threads has two problems: for each endpoint
609
there has to be a number of listening threads. If there are many
610
communication endpoints, it may be difficult to set the right number
611
of concurrent threads in the system, as many of the threads
612
may always be waiting at less busy endpoints. Another problem
613
is queuing of the messages, as the server internally does not
614
offer any queue for jobs.
616
Another group of user threads is intended for splitting the
617
queries and processing them in parallel. Let us call these
618
parallel communication threads. These threads are waiting for
619
parallelized tasks, suspended on event semaphores.
621
A single user thread waits for input from the console,
622
like a command to shut the database.
624
Utility threads are a different group of threads which takes
625
care of the buffer pool flushing and other, mainly background
626
operations, in the server.
627
Some of these utility threads always run at a lower than normal
628
priority, so that they are always in background. Some of them
629
may dynamically boost their priority by the pri_adjust function,
630
even to higher than normal priority, if their task becomes urgent.
631
The running of utilities is controlled by high- and low-water marks
632
of urgency. The urgency may be measured by the number of dirty blocks
633
in the buffer pool, in the case of the flush thread, for example.
634
When the high-water mark is exceeded, an utility starts running, until
635
the urgency drops under the low-water mark. Then the utility thread
636
suspend itself to wait for an event. The master thread is
637
responsible of signaling this event when the utility thread is
640
For each individual type of utility, some threads always remain
641
at lower than normal priority. This is because pri_adjust is implemented
642
so that the threads at normal or higher priority control their
643
share of running time by calling sleep. Thus, if the load of the
644
system sudenly drops, these threads cannot necessarily utilize
645
the system fully. The background priority threads make up for this,
646
starting to run when the load drops.
648
When there is no activity in the system, also the master thread
649
suspends itself to wait for an event making
650
the server totally silent. The responsibility to signal this
651
event is on the user thread which again receives a message
654
There is still one complication in our server design. If a
655
background utility thread obtains a resource (e.g., mutex) needed by a user
656
thread, and there is also some other user activity in the system,
657
the user thread may have to wait indefinitely long for the
658
resource, as the OS does not schedule a background thread if
659
there is some other runnable user thread. This problem is called
660
priority inversion in real-time programming.
662
One solution to the priority inversion problem would be to
663
keep record of which thread owns which resource and
664
in the above case boost the priority of the background thread
665
so that it will be scheduled and it can release the resource.
666
This solution is called priority inheritance in real-time programming.
667
A drawback of this solution is that the overhead of acquiring a mutex
668
increases slightly, maybe 0.2 microseconds on a 100 MHz Pentium, because
669
the thread has to call os_thread_get_curr_id.
670
This may be compared to 0.5 microsecond overhead for a mutex lock-unlock
671
pair. Note that the thread
672
cannot store the information in the resource, say mutex, itself,
673
because competing threads could wipe out the information if it is
674
stored before acquiring the mutex, and if it stored afterwards,
675
the information is outdated for the time of one machine instruction,
676
at least. (To be precise, the information could be stored to
677
lock_word in mutex if the machine supports atomic swap.)
679
The above solution with priority inheritance may become actual in the
680
future, but at the moment we plan to implement a more coarse solution,
681
which could be called a global priority inheritance. If a thread
682
has to wait for a long time, say 300 milliseconds, for a resource,
683
we just guess that it may be waiting for a resource owned by a background
684
thread, and boost the the priority of all runnable background threads
685
to the normal level. The background threads then themselves adjust
686
their fixed priority back to background after releasing all resources
687
they had (or, at some fixed points in their program code).
689
What is the performance of the global priority inheritance solution?
690
We may weigh the length of the wait time 300 milliseconds, during
691
which the system processes some other thread
692
to the cost of boosting the priority of each runnable background
693
thread, rescheduling it, and lowering the priority again.
694
On 100 MHz Pentium + NT this overhead may be of the order 100
695
microseconds per thread. So, if the number of runnable background
696
threads is not very big, say < 100, the cost is tolerable.
697
Utility threads probably will access resources used by
698
user threads not very often, so collisions of user threads
699
to preempted utility threads should not happen very often.
701
The thread table contains
702
information of the current status of each thread existing in the system,
703
and also the event semaphores used in suspending the master thread
704
and utility and parallel communication threads when they have nothing to do.
705
The thread table can be seen as an analogue to the process table
706
in a traditional Unix implementation.
708
The thread table is also used in the global priority inheritance
709
scheme. This brings in one additional complication: threads accessing
710
the thread table must have at least normal fixed priority,
711
because the priority inheritance solution does not work if a background
712
thread is preempted while possessing the mutex protecting the thread table.
713
So, if a thread accesses the thread table, its priority has to be
714
boosted at least to normal. This priority requirement can be seen similar to
715
the privileged mode used when processing the kernel calls in traditional
718
/* Thread slot in the thread table */
719
struct srv_slot_struct{
720
os_thread_id_t id; /*!< thread id */
721
os_thread_t handle; /*!< thread handle */
722
unsigned type:3; /*!< thread type: user, utility etc. */
723
unsigned in_use:1; /*!< TRUE if this slot is in use */
724
unsigned suspended:1; /*!< TRUE if the thread is waiting
725
for the event of this slot */
726
ib_time_t suspend_time; /*!< time when the thread was
728
os_event_t event; /*!< event used in suspending the
729
thread when it has nothing to do */
730
que_thr_t* thr; /*!< suspended query thread (only
731
used for MySQL threads) */
734
/* Table for MySQL threads where they will be suspended to wait for locks */
735
UNIV_INTERN srv_slot_t* srv_mysql_table = NULL;
737
UNIV_INTERN os_event_t srv_lock_timeout_thread_event;
739
UNIV_INTERN srv_sys_t* srv_sys = NULL;
741
/* padding to prevent other memory update hotspots from residing on
742
the same memory cache line */
743
UNIV_INTERN byte srv_pad1[64];
744
/* mutex protecting the server, trx structs, query threads, and lock table */
745
UNIV_INTERN mutex_t* kernel_mutex_temp;
746
/* padding to prevent other memory update hotspots from residing on
747
the same memory cache line */
748
UNIV_INTERN byte srv_pad2[64];
751
/* The following three values measure the urgency of the jobs of
752
buffer, version, and insert threads. They may vary from 0 - 1000.
753
The server mutex protects all these variables. The low-water values
754
tell that the server can acquiesce the utility when the value
755
drops below this low-water mark. */
757
static ulint srv_meter[SRV_MASTER + 1];
758
static ulint srv_meter_low_water[SRV_MASTER + 1];
759
static ulint srv_meter_high_water[SRV_MASTER + 1];
760
static ulint srv_meter_high_water2[SRV_MASTER + 1];
761
static ulint srv_meter_foreground[SRV_MASTER + 1];
764
/***********************************************************************
765
Prints counters for work done by srv_master_thread. */
768
srv_print_master_thread_info(
769
/*=========================*/
770
FILE *file) /* in: output stream */
772
fprintf(file, "srv_master_thread loops: %lu 1_second, %lu sleeps, "
773
"%lu 10_second, %lu background, %lu flush\n",
774
srv_main_1_second_loops, srv_main_sleeps,
775
srv_main_10_second_loops, srv_main_background_loops,
776
srv_main_flush_loops);
777
fprintf(file, "srv_master_thread log flush and writes: %lu\n",
778
srv_log_writes_and_flush);
781
/* The following values give info about the activity going on in
782
the database. They are protected by the server mutex. The arrays
783
are indexed by the type of the thread. */
785
UNIV_INTERN ulint srv_n_threads_active[SRV_MASTER + 1];
786
UNIV_INTERN ulint srv_n_threads[SRV_MASTER + 1];
788
/*********************************************************************//**
789
Sets the info describing an i/o thread current state. */
792
srv_set_io_thread_op_info(
793
/*======================*/
794
ulint i, /*!< in: the 'segment' of the i/o thread */
795
const char* str) /*!< in: constant char string describing the
798
ut_a(i < SRV_MAX_N_IO_THREADS);
800
srv_io_thread_op_info[i] = str;
803
/*********************************************************************//**
804
Accessor function to get pointer to n'th slot in the server thread
806
@return pointer to the slot */
809
srv_table_get_nth_slot(
810
/*===================*/
811
ulint index) /*!< in: index of the slot */
813
ut_a(index < OS_THREAD_MAX_N);
815
return(srv_sys->threads + index);
818
/*********************************************************************//**
819
Gets the number of threads in the system.
820
@return sum of srv_n_threads[] */
823
srv_get_n_threads(void)
824
/*===================*/
829
mutex_enter(&kernel_mutex);
831
for (i = SRV_COM; i < SRV_MASTER + 1; i++) {
833
n_threads += srv_n_threads[i];
836
mutex_exit(&kernel_mutex);
841
/*********************************************************************//**
842
Reserves a slot in the thread table for the current thread. Also creates the
843
thread local storage struct for the current thread. NOTE! The server mutex
844
has to be reserved by the caller!
845
@return reserved slot index */
848
srv_table_reserve_slot(
849
/*===================*/
850
enum srv_thread_type type) /*!< in: type of the thread */
856
ut_a(type <= SRV_MASTER);
859
slot = srv_table_get_nth_slot(i);
861
while (slot->in_use) {
863
slot = srv_table_get_nth_slot(i);
866
ut_a(slot->in_use == FALSE);
869
slot->suspended = FALSE;
871
slot->id = os_thread_get_curr_id();
872
slot->handle = os_thread_get_curr();
876
thr_local_set_slot_no(os_thread_get_curr_id(), i);
881
/*********************************************************************//**
882
Suspends the calling thread to wait for the event in its thread slot.
883
NOTE! The server mutex has to be reserved by the caller!
884
@return event for the calling thread to wait */
887
srv_suspend_thread(void)
888
/*====================*/
893
enum srv_thread_type type;
895
ut_ad(mutex_own(&kernel_mutex));
897
slot_no = thr_local_get_slot_no(os_thread_get_curr_id());
899
if (srv_print_thread_releases) {
901
"Suspending thread %lu to slot %lu\n",
902
(ulong) os_thread_get_curr_id(), (ulong) slot_no);
905
slot = srv_table_get_nth_slot(slot_no);
909
ut_ad(type >= SRV_WORKER);
910
ut_ad(type <= SRV_MASTER);
914
slot->suspended = TRUE;
916
ut_ad(srv_n_threads_active[type] > 0);
918
srv_n_threads_active[type]--;
920
os_event_reset(event);
925
/*********************************************************************//**
926
Releases threads of the type given from suspension in the thread table.
927
NOTE! The server mutex has to be reserved by the caller!
928
@return number of threads released: this may be less than n if not
929
enough threads were suspended at the moment */
934
enum srv_thread_type type, /*!< in: thread type */
935
ulint n) /*!< in: number of threads to release */
941
ut_ad(type >= SRV_WORKER);
942
ut_ad(type <= SRV_MASTER);
944
ut_ad(mutex_own(&kernel_mutex));
946
for (i = 0; i < OS_THREAD_MAX_N; i++) {
948
slot = srv_table_get_nth_slot(i);
950
if (slot->in_use && slot->type == type && slot->suspended) {
952
slot->suspended = FALSE;
954
srv_n_threads_active[type]++;
956
os_event_set(slot->event);
958
if (srv_print_thread_releases) {
960
"Releasing thread %lu type %lu"
962
(ulong) slot->id, (ulong) type,
977
/*********************************************************************//**
978
Returns the calling thread type.
979
@return SRV_COM, ... */
982
srv_get_thread_type(void)
983
/*=====================*/
987
enum srv_thread_type type;
989
mutex_enter(&kernel_mutex);
991
slot_no = thr_local_get_slot_no(os_thread_get_curr_id());
993
slot = srv_table_get_nth_slot(slot_no);
997
ut_ad(type >= SRV_WORKER);
998
ut_ad(type <= SRV_MASTER);
1000
mutex_exit(&kernel_mutex);
1005
/*********************************************************************//**
1006
Initializes the server. */
1012
srv_conc_slot_t* conc_slot;
1016
srv_sys = mem_alloc(sizeof(srv_sys_t));
1018
kernel_mutex_temp = mem_alloc(sizeof(mutex_t));
1019
mutex_create(kernel_mutex_key, &kernel_mutex, SYNC_KERNEL);
1021
mutex_create(srv_innodb_monitor_mutex_key,
1022
&srv_innodb_monitor_mutex, SYNC_NO_ORDER_CHECK);
1024
srv_sys->threads = mem_alloc(OS_THREAD_MAX_N * sizeof(srv_slot_t));
1026
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1027
slot = srv_table_get_nth_slot(i);
1028
slot->in_use = FALSE;
1029
slot->type=0; /* Avoid purify errors */
1030
slot->event = os_event_create(NULL);
1034
srv_mysql_table = mem_alloc(OS_THREAD_MAX_N * sizeof(srv_slot_t));
1036
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1037
slot = srv_mysql_table + i;
1038
slot->in_use = FALSE;
1040
slot->event = os_event_create(NULL);
1044
srv_lock_timeout_thread_event = os_event_create(NULL);
1046
for (i = 0; i < SRV_MASTER + 1; i++) {
1047
srv_n_threads_active[i] = 0;
1048
srv_n_threads[i] = 0;
1051
srv_meter_low_water[i] = 50;
1052
srv_meter_high_water[i] = 100;
1053
srv_meter_high_water2[i] = 200;
1054
srv_meter_foreground[i] = 250;
1058
UT_LIST_INIT(srv_sys->tasks);
1060
/* Create dummy indexes for infimum and supremum records */
1064
/* Init the server concurrency restriction data structures */
1066
os_fast_mutex_init(&srv_conc_mutex);
1068
UT_LIST_INIT(srv_conc_queue);
1070
srv_conc_slots = mem_alloc(OS_THREAD_MAX_N * sizeof(srv_conc_slot_t));
1072
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1073
conc_slot = srv_conc_slots + i;
1074
conc_slot->reserved = FALSE;
1075
conc_slot->event = os_event_create(NULL);
1076
ut_a(conc_slot->event);
1079
/* Initialize some INFORMATION SCHEMA internal structures */
1080
trx_i_s_cache_init(trx_i_s_cache);
1083
/*********************************************************************//**
1084
Frees the data structures created in srv_init(). */
1090
os_fast_mutex_free(&srv_conc_mutex);
1091
mem_free(srv_conc_slots);
1092
srv_conc_slots = NULL;
1094
mem_free(srv_sys->threads);
1098
mem_free(kernel_mutex_temp);
1099
kernel_mutex_temp = NULL;
1100
mem_free(srv_mysql_table);
1101
srv_mysql_table = NULL;
1103
trx_i_s_cache_free(trx_i_s_cache);
1106
/*********************************************************************//**
1107
Initializes the synchronization primitives, memory system, and the thread
1111
srv_general_init(void)
1112
/*==================*/
1115
/* Reset the system variables in the recovery module. */
1116
recv_sys_var_init();
1119
mem_init(srv_mem_pool_size);
1123
/*======================= InnoDB Server FIFO queue =======================*/
1125
/* Maximum allowable purge history length. <=0 means 'infinite'. */
1126
UNIV_INTERN ulong srv_max_purge_lag = 0;
1128
/*********************************************************************//**
1129
Puts an OS thread to wait if there are too many concurrent threads
1130
(>= srv_thread_concurrency) inside InnoDB. The threads wait in a FIFO queue. */
1133
srv_conc_enter_innodb(
1134
/*==================*/
1135
trx_t* trx) /*!< in: transaction object associated with the
1138
ibool has_slept = FALSE;
1139
srv_conc_slot_t* slot = NULL;
1142
if (trx->mysql_thd != NULL
1143
&& thd_is_replication_slave_thread(trx->mysql_thd)) {
1145
UT_WAIT_FOR(srv_conc_n_threads
1146
< (lint)srv_thread_concurrency,
1147
srv_replication_delay * 1000);
1152
/* If trx has 'free tickets' to enter the engine left, then use one
1155
if (trx->n_tickets_to_enter_innodb > 0) {
1156
trx->n_tickets_to_enter_innodb--;
1161
os_fast_mutex_lock(&srv_conc_mutex);
1163
if (trx->declared_to_be_inside_innodb) {
1164
ut_print_timestamp(stderr);
1165
fputs(" InnoDB: Error: trying to declare trx"
1166
" to enter InnoDB, but\n"
1167
"InnoDB: it already is declared.\n", stderr);
1168
trx_print(stderr, trx, 0);
1170
os_fast_mutex_unlock(&srv_conc_mutex);
1175
ut_ad(srv_conc_n_threads >= 0);
1177
if (srv_conc_n_threads < (lint)srv_thread_concurrency) {
1179
srv_conc_n_threads++;
1180
trx->declared_to_be_inside_innodb = TRUE;
1181
trx->n_tickets_to_enter_innodb = SRV_FREE_TICKETS_TO_ENTER;
1183
os_fast_mutex_unlock(&srv_conc_mutex);
1188
/* If the transaction is not holding resources, let it sleep
1189
for SRV_THREAD_SLEEP_DELAY microseconds, and try again then */
1191
if (!has_slept && !trx->has_search_latch
1192
&& NULL == UT_LIST_GET_FIRST(trx->trx_locks)) {
1194
has_slept = TRUE; /* We let it sleep only once to avoid
1197
srv_conc_n_waiting_threads++;
1199
os_fast_mutex_unlock(&srv_conc_mutex);
1201
trx->op_info = "sleeping before joining InnoDB queue";
1203
/* Peter Zaitsev suggested that we take the sleep away
1204
altogether. But the sleep may be good in pathological
1205
situations of lots of thread switches. Simply put some
1206
threads aside for a while to reduce the number of thread
1208
if (SRV_THREAD_SLEEP_DELAY > 0) {
1209
os_thread_sleep(SRV_THREAD_SLEEP_DELAY);
1214
os_fast_mutex_lock(&srv_conc_mutex);
1216
srv_conc_n_waiting_threads--;
1221
/* Too many threads inside: put the current thread to a queue */
1223
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1224
slot = srv_conc_slots + i;
1226
if (!slot->reserved) {
1232
if (i == OS_THREAD_MAX_N) {
1233
/* Could not find a free wait slot, we must let the
1236
srv_conc_n_threads++;
1237
trx->declared_to_be_inside_innodb = TRUE;
1238
trx->n_tickets_to_enter_innodb = 0;
1240
os_fast_mutex_unlock(&srv_conc_mutex);
1245
/* Release possible search system latch this thread has */
1246
if (trx->has_search_latch) {
1247
trx_search_latch_release_if_reserved(trx);
1250
/* Add to the queue */
1251
slot->reserved = TRUE;
1252
slot->wait_ended = FALSE;
1254
UT_LIST_ADD_LAST(srv_conc_queue, srv_conc_queue, slot);
1256
os_event_reset(slot->event);
1258
srv_conc_n_waiting_threads++;
1260
os_fast_mutex_unlock(&srv_conc_mutex);
1262
/* Go to wait for the event; when a thread leaves InnoDB it will
1263
release this thread */
1265
trx->op_info = "waiting in InnoDB queue";
1267
os_event_wait(slot->event);
1271
os_fast_mutex_lock(&srv_conc_mutex);
1273
srv_conc_n_waiting_threads--;
1275
/* NOTE that the thread which released this thread already
1276
incremented the thread counter on behalf of this thread */
1278
slot->reserved = FALSE;
1280
UT_LIST_REMOVE(srv_conc_queue, srv_conc_queue, slot);
1282
trx->declared_to_be_inside_innodb = TRUE;
1283
trx->n_tickets_to_enter_innodb = SRV_FREE_TICKETS_TO_ENTER;
1285
os_fast_mutex_unlock(&srv_conc_mutex);
1288
/*********************************************************************//**
1289
This lets a thread enter InnoDB regardless of the number of threads inside
1290
InnoDB. This must be called when a thread ends a lock wait. */
1293
srv_conc_force_enter_innodb(
1294
/*========================*/
1295
trx_t* trx) /*!< in: transaction object associated with the
1298
if (UNIV_LIKELY(!srv_thread_concurrency)) {
1303
ut_ad(srv_conc_n_threads >= 0);
1305
os_fast_mutex_lock(&srv_conc_mutex);
1307
srv_conc_n_threads++;
1308
trx->declared_to_be_inside_innodb = TRUE;
1309
trx->n_tickets_to_enter_innodb = 1;
1311
os_fast_mutex_unlock(&srv_conc_mutex);
1314
/*********************************************************************//**
1315
This must be called when a thread exits InnoDB in a lock wait or at the
1316
end of an SQL statement. */
1319
srv_conc_force_exit_innodb(
1320
/*=======================*/
1321
trx_t* trx) /*!< in: transaction object associated with the
1324
srv_conc_slot_t* slot = NULL;
1326
if (trx->mysql_thd != NULL
1327
&& thd_is_replication_slave_thread(trx->mysql_thd)) {
1332
if (trx->declared_to_be_inside_innodb == FALSE) {
1337
os_fast_mutex_lock(&srv_conc_mutex);
1339
ut_ad(srv_conc_n_threads > 0);
1340
srv_conc_n_threads--;
1341
trx->declared_to_be_inside_innodb = FALSE;
1342
trx->n_tickets_to_enter_innodb = 0;
1344
if (srv_conc_n_threads < (lint)srv_thread_concurrency) {
1345
/* Look for a slot where a thread is waiting and no other
1346
thread has yet released the thread */
1348
slot = UT_LIST_GET_FIRST(srv_conc_queue);
1350
while (slot && slot->wait_ended == TRUE) {
1351
slot = UT_LIST_GET_NEXT(srv_conc_queue, slot);
1355
slot->wait_ended = TRUE;
1357
/* We increment the count on behalf of the released
1360
srv_conc_n_threads++;
1364
os_fast_mutex_unlock(&srv_conc_mutex);
1367
os_event_set(slot->event);
1371
/*********************************************************************//**
1372
This must be called when a thread exits InnoDB. */
1375
srv_conc_exit_innodb(
1376
/*=================*/
1377
trx_t* trx) /*!< in: transaction object associated with the
1380
if (trx->n_tickets_to_enter_innodb > 0) {
1381
/* We will pretend the thread is still inside InnoDB though it
1382
now leaves the InnoDB engine. In this way we save
1383
a lot of semaphore operations. srv_conc_force_exit_innodb is
1384
used to declare the thread definitely outside InnoDB. It
1385
should be called when there is a lock wait or an SQL statement
1391
srv_conc_force_exit_innodb(trx);
1394
/*========================================================================*/
1396
/*********************************************************************//**
1397
Normalizes init parameter values to use units we use inside InnoDB.
1398
@return DB_SUCCESS or error code */
1401
srv_normalize_init_values(void)
1402
/*===========================*/
1407
n = srv_n_data_files;
1409
for (i = 0; i < n; i++) {
1410
srv_data_file_sizes[i] = srv_data_file_sizes[i]
1411
* ((1024 * 1024) / UNIV_PAGE_SIZE);
1414
srv_last_file_size_max = srv_last_file_size_max
1415
* ((1024 * 1024) / UNIV_PAGE_SIZE);
1417
srv_log_file_size = srv_log_file_size / UNIV_PAGE_SIZE;
1419
srv_log_buffer_size = srv_log_buffer_size / UNIV_PAGE_SIZE;
1421
srv_lock_table_size = 5 * (srv_buf_pool_size / UNIV_PAGE_SIZE);
1426
/*********************************************************************//**
1427
Boots the InnoDB server.
1428
@return DB_SUCCESS or error code */
1436
/* Transform the init parameter values given by MySQL to
1437
use units we use inside InnoDB: */
1439
err = srv_normalize_init_values();
1441
if (err != DB_SUCCESS) {
1445
/* Initialize synchronization primitives, memory management, and thread
1450
/* Initialize this module */
1457
/*********************************************************************//**
1458
Reserves a slot in the thread table for the current MySQL OS thread.
1459
NOTE! The kernel mutex has to be reserved by the caller!
1460
@return reserved slot */
1463
srv_table_reserve_slot_for_mysql(void)
1464
/*==================================*/
1469
ut_ad(mutex_own(&kernel_mutex));
1472
slot = srv_mysql_table + i;
1474
while (slot->in_use) {
1477
if (i >= OS_THREAD_MAX_N) {
1479
ut_print_timestamp(stderr);
1482
" InnoDB: There appear to be %lu MySQL"
1483
" threads currently waiting\n"
1484
"InnoDB: inside InnoDB, which is the"
1485
" upper limit. Cannot continue operation.\n"
1486
"InnoDB: We intentionally generate"
1487
" a seg fault to print a stack trace\n"
1488
"InnoDB: on Linux. But first we print"
1489
" a list of waiting threads.\n", (ulong) i);
1491
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1493
slot = srv_mysql_table + i;
1496
"Slot %lu: thread id %lu, type %lu,"
1497
" in use %lu, susp %lu, time %lu\n",
1499
(ulong) os_thread_pf(slot->id),
1501
(ulong) slot->in_use,
1502
(ulong) slot->suspended,
1503
(ulong) difftime(ut_time(),
1504
slot->suspend_time));
1510
slot = srv_mysql_table + i;
1513
ut_a(slot->in_use == FALSE);
1515
slot->in_use = TRUE;
1516
slot->id = os_thread_get_curr_id();
1517
slot->handle = os_thread_get_curr();
1522
/***************************************************************//**
1523
Puts a MySQL OS thread to wait for a lock to be released. If an error
1524
occurs during the wait trx->error_state associated with thr is
1525
!= DB_SUCCESS when we return. DB_LOCK_WAIT_TIMEOUT and DB_DEADLOCK
1526
are possible errors. DB_DEADLOCK is returned if selective deadlock
1527
resolution chose this transaction as a victim. */
1530
srv_suspend_mysql_thread(
1531
/*=====================*/
1532
que_thr_t* thr) /*!< in: query thread associated with the MySQL
1539
ulint had_dict_lock;
1540
ibool was_declared_inside_innodb = FALSE;
1541
ib_int64_t start_time = 0;
1542
ib_int64_t finish_time;
1546
ulong lock_wait_timeout;
1548
ut_ad(!mutex_own(&kernel_mutex));
1550
trx = thr_get_trx(thr);
1552
os_event_set(srv_lock_timeout_thread_event);
1554
mutex_enter(&kernel_mutex);
1556
trx->error_state = DB_SUCCESS;
1558
if (thr->state == QUE_THR_RUNNING) {
1560
ut_ad(thr->is_active == TRUE);
1562
/* The lock has already been released or this transaction
1563
was chosen as a deadlock victim: no need to suspend */
1565
if (trx->was_chosen_as_deadlock_victim) {
1567
trx->error_state = DB_DEADLOCK;
1568
trx->was_chosen_as_deadlock_victim = FALSE;
1571
mutex_exit(&kernel_mutex);
1576
ut_ad(thr->is_active == FALSE);
1578
slot = srv_table_reserve_slot_for_mysql();
1580
event = slot->event;
1584
os_event_reset(event);
1586
slot->suspend_time = ut_time();
1588
if (thr->lock_state == QUE_THR_LOCK_ROW) {
1589
srv_n_lock_wait_count++;
1590
srv_n_lock_wait_current_count++;
1592
if (ut_usectime(&sec, &ms) == -1) {
1595
start_time = (ib_int64_t) sec * 1000000 + ms;
1598
/* Wake the lock timeout monitor thread, if it is suspended */
1600
os_event_set(srv_lock_timeout_thread_event);
1602
mutex_exit(&kernel_mutex);
1604
if (trx->declared_to_be_inside_innodb) {
1606
was_declared_inside_innodb = TRUE;
1608
/* We must declare this OS thread to exit InnoDB, since a
1609
possible other thread holding a lock which this thread waits
1610
for must be allowed to enter, sooner or later */
1612
srv_conc_force_exit_innodb(trx);
1615
had_dict_lock = trx->dict_operation_lock_mode;
1617
switch (had_dict_lock) {
1619
/* Release foreign key check latch */
1620
row_mysql_unfreeze_data_dictionary(trx);
1623
/* There should never be a lock wait when the
1624
dictionary latch is reserved in X mode. Dictionary
1625
transactions should only acquire locks on dictionary
1626
tables, not other tables. All access to dictionary
1627
tables should be covered by dictionary
1629
ut_print_timestamp(stderr);
1630
fputs(" InnoDB: Error: dict X latch held in "
1631
"srv_suspend_mysql_thread\n", stderr);
1632
/* This should never occur. This incorrect handling
1633
was added in the early development of
1634
ha_innobase::add_index() in InnoDB Plugin 1.0. */
1635
/* Release fast index creation latch */
1636
row_mysql_unlock_data_dictionary(trx);
1640
ut_a(trx->dict_operation_lock_mode == 0);
1642
/* Suspend this thread and wait for the event. */
1644
os_event_wait(event);
1646
/* After resuming, reacquire the data dictionary latch if
1649
switch (had_dict_lock) {
1651
row_mysql_freeze_data_dictionary(trx);
1654
/* This should never occur. This incorrect handling
1655
was added in the early development of
1656
ha_innobase::add_index() in InnoDB Plugin 1.0. */
1657
row_mysql_lock_data_dictionary(trx);
1661
if (was_declared_inside_innodb) {
1663
/* Return back inside InnoDB */
1665
srv_conc_force_enter_innodb(trx);
1668
mutex_enter(&kernel_mutex);
1670
/* Release the slot for others to use */
1672
slot->in_use = FALSE;
1674
wait_time = ut_difftime(ut_time(), slot->suspend_time);
1676
if (thr->lock_state == QUE_THR_LOCK_ROW) {
1677
if (ut_usectime(&sec, &ms) == -1) {
1680
finish_time = (ib_int64_t) sec * 1000000 + ms;
1683
diff_time = (ulint) (finish_time - start_time);
1685
srv_n_lock_wait_current_count--;
1686
srv_n_lock_wait_time = srv_n_lock_wait_time + diff_time;
1687
if (diff_time > srv_n_lock_max_wait_time &&
1688
/* only update the variable if we successfully
1689
retrieved the start and finish times. See Bug#36819. */
1690
start_time != -1 && finish_time != -1) {
1691
srv_n_lock_max_wait_time = diff_time;
1694
/* Record the lock wait time for this thread */
1695
thd_set_lock_wait_time(trx->mysql_thd, diff_time);
1698
if (trx->was_chosen_as_deadlock_victim) {
1700
trx->error_state = DB_DEADLOCK;
1701
trx->was_chosen_as_deadlock_victim = FALSE;
1704
mutex_exit(&kernel_mutex);
1706
/* InnoDB system transactions (such as the purge, and
1707
incomplete transactions that are being rolled back after crash
1708
recovery) will use the global value of
1709
innodb_lock_wait_timeout, because trx->mysql_thd == NULL. */
1710
lock_wait_timeout = thd_lock_wait_timeout(trx->mysql_thd);
1712
if (lock_wait_timeout < 100000000
1713
&& wait_time > (double) lock_wait_timeout) {
1715
trx->error_state = DB_LOCK_WAIT_TIMEOUT;
1718
if (trx_is_interrupted(trx)) {
1720
trx->error_state = DB_INTERRUPTED;
1724
/********************************************************************//**
1725
Releases a MySQL OS thread waiting for a lock to be released, if the
1726
thread is already suspended. */
1729
srv_release_mysql_thread_if_suspended(
1730
/*==================================*/
1731
que_thr_t* thr) /*!< in: query thread associated with the
1737
ut_ad(mutex_own(&kernel_mutex));
1739
for (i = 0; i < OS_THREAD_MAX_N; i++) {
1741
slot = srv_mysql_table + i;
1743
if (slot->in_use && slot->thr == thr) {
1746
os_event_set(slot->event);
1755
/******************************************************************//**
1756
Refreshes the values used to calculate per-second averages. */
1759
srv_refresh_innodb_monitor_stats(void)
1760
/*==================================*/
1762
mutex_enter(&srv_innodb_monitor_mutex);
1764
srv_last_monitor_time = time(NULL);
1766
os_aio_refresh_stats();
1768
btr_cur_n_sea_old = btr_cur_n_sea;
1769
btr_cur_n_non_sea_old = btr_cur_n_non_sea;
1771
log_refresh_stats();
1773
buf_refresh_io_stats_all();
1775
srv_n_rows_inserted_old = srv_n_rows_inserted;
1776
srv_n_rows_updated_old = srv_n_rows_updated;
1777
srv_n_rows_deleted_old = srv_n_rows_deleted;
1778
srv_n_rows_read_old = srv_n_rows_read;
1780
mutex_exit(&srv_innodb_monitor_mutex);
1783
/******************************************************************//**
1784
Outputs to a file the output of the InnoDB Monitor.
1785
@return FALSE if not all information printed
1786
due to failure to obtain necessary mutex */
1789
srv_printf_innodb_monitor(
1790
/*======================*/
1791
FILE* file, /*!< in: output stream */
1792
ibool nowait, /*!< in: whether to wait for kernel mutex */
1793
ulint* trx_start, /*!< out: file position of the start of
1794
the list of active transactions */
1795
ulint* trx_end) /*!< out: file position of the end of
1796
the list of active transactions */
1798
double time_elapsed;
1799
time_t current_time;
1803
mutex_enter(&srv_innodb_monitor_mutex);
1805
current_time = time(NULL);
1807
/* We add 0.001 seconds to time_elapsed to prevent division
1808
by zero if two users happen to call SHOW INNODB STATUS at the same
1811
time_elapsed = difftime(current_time, srv_last_monitor_time)
1814
srv_last_monitor_time = time(NULL);
1816
fputs("\n=====================================\n", file);
1818
ut_print_timestamp(file);
1820
" INNODB MONITOR OUTPUT\n"
1821
"=====================================\n"
1822
"Per second averages calculated from the last %lu seconds\n",
1823
(ulong)time_elapsed);
1825
fputs("-----------------\n"
1826
"BACKGROUND THREAD\n"
1827
"-----------------\n", file);
1828
srv_print_master_thread_info(file);
1830
fputs("----------\n"
1832
"----------\n", file);
1835
/* Conceptually, srv_innodb_monitor_mutex has a very high latching
1836
order level in sync0sync.h, while dict_foreign_err_mutex has a very
1837
low level 135. Therefore we can reserve the latter mutex here without
1838
a danger of a deadlock of threads. */
1840
mutex_enter(&dict_foreign_err_mutex);
1842
if (ftell(dict_foreign_err_file) != 0L) {
1843
fputs("------------------------\n"
1844
"LATEST FOREIGN KEY ERROR\n"
1845
"------------------------\n", file);
1846
ut_copy_file(file, dict_foreign_err_file);
1849
mutex_exit(&dict_foreign_err_mutex);
1851
/* Only if lock_print_info_summary proceeds correctly,
1852
before we call the lock_print_info_all_transactions
1853
to print all the lock information. */
1854
ret = lock_print_info_summary(file, nowait);
1858
long t = ftell(file);
1860
*trx_start = ULINT_UNDEFINED;
1862
*trx_start = (ulint) t;
1865
lock_print_info_all_transactions(file);
1867
long t = ftell(file);
1869
*trx_end = ULINT_UNDEFINED;
1871
*trx_end = (ulint) t;
1878
"--------\n", file);
1881
fputs("-------------------------------------\n"
1882
"INSERT BUFFER AND ADAPTIVE HASH INDEX\n"
1883
"-------------------------------------\n", file);
1886
ha_print_info(file, btr_search_sys->hash_index);
1889
"%.2f hash searches/s, %.2f non-hash searches/s\n",
1890
(btr_cur_n_sea - btr_cur_n_sea_old)
1892
(btr_cur_n_non_sea - btr_cur_n_non_sea_old)
1894
btr_cur_n_sea_old = btr_cur_n_sea;
1895
btr_cur_n_non_sea_old = btr_cur_n_non_sea;
1902
fputs("----------------------\n"
1903
"BUFFER POOL AND MEMORY\n"
1904
"----------------------\n", file);
1906
"Total memory allocated " ULINTPF
1907
"; in additional pool allocated " ULINTPF "\n",
1908
ut_total_allocated_memory,
1909
mem_pool_get_reserved(mem_comm_pool));
1910
fprintf(file, "Dictionary memory allocated " ULINTPF "\n",
1915
fputs("--------------\n"
1917
"--------------\n", file);
1918
fprintf(file, "%ld queries inside InnoDB, %lu queries in queue\n",
1919
(long) srv_conc_n_threads,
1920
(ulong) srv_conc_n_waiting_threads);
1922
fprintf(file, "%lu read views open inside InnoDB\n",
1923
UT_LIST_GET_LEN(trx_sys->view_list));
1925
n_reserved = fil_space_get_n_reserved_extents(0);
1926
if (n_reserved > 0) {
1928
"%lu tablespace extents now reserved for"
1929
" B-tree split operations\n",
1930
(ulong) n_reserved);
1934
fprintf(file, "Main thread process no. %lu, id %lu, state: %s\n",
1935
(ulong) srv_main_thread_process_no,
1936
(ulong) srv_main_thread_id,
1937
srv_main_thread_op_info);
1939
fprintf(file, "Main thread id %lu, state: %s\n",
1940
(ulong) srv_main_thread_id,
1941
srv_main_thread_op_info);
1944
"Number of rows inserted " ULINTPF
1945
", updated " ULINTPF ", deleted " ULINTPF
1946
", read " ULINTPF "\n",
1947
srv_n_rows_inserted,
1952
"%.2f inserts/s, %.2f updates/s,"
1953
" %.2f deletes/s, %.2f reads/s\n",
1954
(srv_n_rows_inserted - srv_n_rows_inserted_old)
1956
(srv_n_rows_updated - srv_n_rows_updated_old)
1958
(srv_n_rows_deleted - srv_n_rows_deleted_old)
1960
(srv_n_rows_read - srv_n_rows_read_old)
1963
srv_n_rows_inserted_old = srv_n_rows_inserted;
1964
srv_n_rows_updated_old = srv_n_rows_updated;
1965
srv_n_rows_deleted_old = srv_n_rows_deleted;
1966
srv_n_rows_read_old = srv_n_rows_read;
1968
fputs("----------------------------\n"
1969
"END OF INNODB MONITOR OUTPUT\n"
1970
"============================\n", file);
1971
mutex_exit(&srv_innodb_monitor_mutex);
1977
/******************************************************************//**
1978
Function to pass InnoDB status variables to MySQL */
1981
srv_export_innodb_status(void)
1982
/*==========================*/
1984
buf_pool_stat_t stat;
1987
ulint flush_list_len;
1989
buf_get_total_stat(&stat);
1990
buf_get_total_list_len(&LRU_len, &free_len, &flush_list_len);
1992
mutex_enter(&srv_innodb_monitor_mutex);
1994
export_vars.innodb_data_pending_reads
1995
= os_n_pending_reads;
1996
export_vars.innodb_data_pending_writes
1997
= os_n_pending_writes;
1998
export_vars.innodb_data_pending_fsyncs
1999
= fil_n_pending_log_flushes
2000
+ fil_n_pending_tablespace_flushes;
2001
export_vars.innodb_data_fsyncs = os_n_fsyncs;
2002
export_vars.innodb_data_read = srv_data_read;
2003
export_vars.innodb_data_reads = os_n_file_reads;
2004
export_vars.innodb_data_writes = os_n_file_writes;
2005
export_vars.innodb_data_written = srv_data_written;
2006
export_vars.innodb_buffer_pool_read_requests = stat.n_page_gets;
2007
export_vars.innodb_buffer_pool_write_requests
2008
= srv_buf_pool_write_requests;
2009
export_vars.innodb_buffer_pool_wait_free = srv_buf_pool_wait_free;
2010
export_vars.innodb_buffer_pool_pages_flushed = srv_buf_pool_flushed;
2011
export_vars.innodb_buffer_pool_reads = srv_buf_pool_reads;
2012
export_vars.innodb_buffer_pool_read_ahead
2013
= stat.n_ra_pages_read;
2014
export_vars.innodb_buffer_pool_read_ahead_evicted
2015
= stat.n_ra_pages_evicted;
2016
export_vars.innodb_buffer_pool_pages_data = LRU_len;
2017
export_vars.innodb_buffer_pool_pages_dirty = flush_list_len;
2018
export_vars.innodb_buffer_pool_pages_free = free_len;
2020
export_vars.innodb_buffer_pool_pages_latched
2021
= buf_get_latched_pages_number();
2022
#endif /* UNIV_DEBUG */
2023
export_vars.innodb_buffer_pool_pages_total = buf_pool_get_n_pages();
2025
export_vars.innodb_buffer_pool_pages_misc
2026
= buf_pool_get_n_pages() - LRU_len - free_len;
2027
#ifdef HAVE_ATOMIC_BUILTINS
2028
export_vars.innodb_have_atomic_builtins = 1;
2030
export_vars.innodb_have_atomic_builtins = 0;
2032
export_vars.innodb_page_size = UNIV_PAGE_SIZE;
2033
export_vars.innodb_log_waits = srv_log_waits;
2034
export_vars.innodb_os_log_written = srv_os_log_written;
2035
export_vars.innodb_os_log_fsyncs = fil_n_log_flushes;
2036
export_vars.innodb_os_log_pending_fsyncs = fil_n_pending_log_flushes;
2037
export_vars.innodb_os_log_pending_writes = srv_os_log_pending_writes;
2038
export_vars.innodb_log_write_requests = srv_log_write_requests;
2039
export_vars.innodb_log_writes = srv_log_writes;
2040
export_vars.innodb_dblwr_pages_written = srv_dblwr_pages_written;
2041
export_vars.innodb_dblwr_writes = srv_dblwr_writes;
2042
export_vars.innodb_pages_created = stat.n_pages_created;
2043
export_vars.innodb_pages_read = stat.n_pages_read;
2044
export_vars.innodb_pages_written = stat.n_pages_written;
2045
export_vars.innodb_row_lock_waits = srv_n_lock_wait_count;
2046
export_vars.innodb_row_lock_current_waits
2047
= srv_n_lock_wait_current_count;
2048
export_vars.innodb_row_lock_time = srv_n_lock_wait_time / 1000;
2049
if (srv_n_lock_wait_count > 0) {
2050
export_vars.innodb_row_lock_time_avg = (ulint)
2051
(srv_n_lock_wait_time / 1000 / srv_n_lock_wait_count);
2053
export_vars.innodb_row_lock_time_avg = 0;
2055
export_vars.innodb_row_lock_time_max
2056
= srv_n_lock_max_wait_time / 1000;
2057
export_vars.innodb_rows_read = srv_n_rows_read;
2058
export_vars.innodb_rows_inserted = srv_n_rows_inserted;
2059
export_vars.innodb_rows_updated = srv_n_rows_updated;
2060
export_vars.innodb_rows_deleted = srv_n_rows_deleted;
2061
export_vars.innodb_truncated_status_writes = srv_truncated_status_writes;
2063
mutex_exit(&srv_innodb_monitor_mutex);
2066
/*********************************************************************//**
2067
A thread which prints the info output by various InnoDB monitors.
2068
@return a dummy parameter */
2073
void* arg __attribute__((unused)))
2074
/*!< in: a dummy parameter required by
2077
double time_elapsed;
2078
time_t current_time;
2079
time_t last_table_monitor_time;
2080
time_t last_tablespace_monitor_time;
2081
time_t last_monitor_time;
2082
ulint mutex_skipped;
2083
ibool last_srv_print_monitor;
2085
#ifdef UNIV_DEBUG_THREAD_CREATION
2086
fprintf(stderr, "Lock timeout thread starts, id %lu\n",
2087
os_thread_pf(os_thread_get_curr_id()));
2090
#ifdef UNIV_PFS_THREAD
2091
pfs_register_thread(srv_monitor_thread_key);
2095
srv_last_monitor_time = time(NULL);
2096
last_table_monitor_time = time(NULL);
2097
last_tablespace_monitor_time = time(NULL);
2098
last_monitor_time = time(NULL);
2100
last_srv_print_monitor = srv_print_innodb_monitor;
2102
srv_monitor_active = TRUE;
2104
/* Wake up every 5 seconds to see if we need to print
2105
monitor information. */
2107
os_thread_sleep(5000000);
2109
current_time = time(NULL);
2111
time_elapsed = difftime(current_time, last_monitor_time);
2113
if (time_elapsed > 15) {
2114
last_monitor_time = time(NULL);
2116
if (srv_print_innodb_monitor) {
2117
/* Reset mutex_skipped counter everytime
2118
srv_print_innodb_monitor changes. This is to
2119
ensure we will not be blocked by kernel_mutex
2120
for short duration information printing,
2121
such as requested by sync_array_print_long_waits() */
2122
if (!last_srv_print_monitor) {
2124
last_srv_print_monitor = TRUE;
2127
if (!srv_printf_innodb_monitor(stderr,
2128
MUTEX_NOWAIT(mutex_skipped),
2132
/* Reset the counter */
2136
last_srv_print_monitor = FALSE;
2140
if (srv_innodb_status) {
2141
mutex_enter(&srv_monitor_file_mutex);
2142
rewind(srv_monitor_file);
2143
if (!srv_printf_innodb_monitor(srv_monitor_file,
2144
MUTEX_NOWAIT(mutex_skipped),
2151
os_file_set_eof(srv_monitor_file);
2152
mutex_exit(&srv_monitor_file_mutex);
2155
if (srv_print_innodb_tablespace_monitor
2156
&& difftime(current_time,
2157
last_tablespace_monitor_time) > 60) {
2158
last_tablespace_monitor_time = time(NULL);
2160
fputs("========================"
2161
"========================\n",
2164
ut_print_timestamp(stderr);
2166
fputs(" INNODB TABLESPACE MONITOR OUTPUT\n"
2167
"========================"
2168
"========================\n",
2172
fputs("Validating tablespace\n", stderr);
2174
fputs("Validation ok\n"
2175
"---------------------------------------\n"
2176
"END OF INNODB TABLESPACE MONITOR OUTPUT\n"
2177
"=======================================\n",
2181
if (srv_print_innodb_table_monitor
2182
&& difftime(current_time, last_table_monitor_time) > 60) {
2184
last_table_monitor_time = time(NULL);
2186
fputs("===========================================\n",
2189
ut_print_timestamp(stderr);
2191
fputs(" INNODB TABLE MONITOR OUTPUT\n"
2192
"===========================================\n",
2196
fputs("-----------------------------------\n"
2197
"END OF INNODB TABLE MONITOR OUTPUT\n"
2198
"==================================\n",
2203
if (srv_shutdown_state >= SRV_SHUTDOWN_CLEANUP) {
2207
if (srv_print_innodb_monitor
2208
|| srv_print_innodb_lock_monitor
2209
|| srv_print_innodb_tablespace_monitor
2210
|| srv_print_innodb_table_monitor) {
2214
srv_monitor_active = FALSE;
2219
srv_monitor_active = FALSE;
2221
/* We count the number of threads in os_thread_exit(). A created
2222
thread should always use that to exit and not use return() to exit. */
2224
os_thread_exit(NULL);
2226
OS_THREAD_DUMMY_RETURN;
2229
/*********************************************************************//**
2230
A thread which wakes up threads whose lock wait may have lasted too long.
2231
@return a dummy parameter */
2234
srv_lock_timeout_thread(
2235
/*====================*/
2236
void* arg __attribute__((unused)))
2237
/* in: a dummy parameter required by
2245
#ifdef UNIV_PFS_THREAD
2246
pfs_register_thread(srv_lock_timeout_thread_key);
2250
/* When someone is waiting for a lock, we wake up every second
2251
and check if a timeout has passed for a lock wait */
2253
os_thread_sleep(1000000);
2255
srv_lock_timeout_active = TRUE;
2257
mutex_enter(&kernel_mutex);
2261
/* Check of all slots if a thread is waiting there, and if it
2262
has exceeded the time limit */
2264
for (i = 0; i < OS_THREAD_MAX_N; i++) {
2266
slot = srv_mysql_table + i;
2270
ulong lock_wait_timeout;
2274
wait_time = ut_difftime(ut_time(), slot->suspend_time);
2276
trx = thr_get_trx(slot->thr);
2277
lock_wait_timeout = thd_lock_wait_timeout(
2280
if (trx_is_interrupted(trx)
2281
|| (lock_wait_timeout < 100000000
2282
&& (wait_time > (double) lock_wait_timeout
2283
|| wait_time < 0))) {
2285
/* Timeout exceeded or a wrap-around in system
2286
time counter: cancel the lock request queued
2287
by the transaction and release possible
2288
other transactions waiting behind; it is
2289
possible that the lock has already been
2290
granted: in that case do nothing */
2292
if (trx->wait_lock) {
2293
lock_cancel_waiting_and_release(
2300
os_event_reset(srv_lock_timeout_thread_event);
2302
mutex_exit(&kernel_mutex);
2304
if (srv_shutdown_state >= SRV_SHUTDOWN_CLEANUP) {
2312
srv_lock_timeout_active = FALSE;
2315
/* The following synchronisation is disabled, since
2316
the InnoDB monitor output is to be updated every 15 seconds. */
2317
os_event_wait(srv_lock_timeout_thread_event);
2322
srv_lock_timeout_active = FALSE;
2324
/* We count the number of threads in os_thread_exit(). A created
2325
thread should always use that to exit and not use return() to exit. */
2327
os_thread_exit(NULL);
2329
OS_THREAD_DUMMY_RETURN;
2332
/*********************************************************************//**
2333
A thread which prints warnings about semaphore waits which have lasted
2334
too long. These can be used to track bugs which cause hangs.
2335
@return a dummy parameter */
2338
srv_error_monitor_thread(
2339
/*=====================*/
2340
void* arg __attribute__((unused)))
2341
/*!< in: a dummy parameter required by
2344
/* number of successive fatal timeouts observed */
2345
ulint fatal_cnt = 0;
2346
ib_uint64_t old_lsn;
2347
ib_uint64_t new_lsn;
2349
old_lsn = srv_start_lsn;
2351
#ifdef UNIV_DEBUG_THREAD_CREATION
2352
fprintf(stderr, "Error monitor thread starts, id %lu\n",
2353
os_thread_pf(os_thread_get_curr_id()));
2356
#ifdef UNIV_PFS_THREAD
2357
pfs_register_thread(srv_error_monitor_thread_key);
2361
srv_error_monitor_active = TRUE;
2363
/* Try to track a strange bug reported by Harald Fuchs and others,
2364
where the lsn seems to decrease at times */
2366
new_lsn = log_get_lsn();
2368
if (new_lsn < old_lsn) {
2369
ut_print_timestamp(stderr);
2371
" InnoDB: Error: old log sequence number %"PRIu64""
2373
"InnoDB: than the new log sequence number %"PRIu64"!\n"
2374
"InnoDB: Please submit a bug report"
2375
" to http://bugs.mysql.com\n",
2381
if (difftime(time(NULL), srv_last_monitor_time) > 60) {
2382
/* We referesh InnoDB Monitor values so that averages are
2383
printed from at most 60 last seconds */
2385
srv_refresh_innodb_monitor_stats();
2388
/* Update the statistics collected for deciding LRU
2390
buf_LRU_stat_update();
2392
/* Update the statistics collected for flush rate policy. */
2393
buf_flush_stat_update();
2395
/* In case mutex_exit is not a memory barrier, it is
2396
theoretically possible some threads are left waiting though
2397
the semaphore is already released. Wake up those threads: */
2399
sync_arr_wake_threads_if_sema_free();
2401
if (sync_array_print_long_waits()) {
2403
if (fatal_cnt > 10) {
2406
"InnoDB: Error: semaphore wait has lasted"
2408
"InnoDB: We intentionally crash the server,"
2409
" because it appears to be hung.\n",
2410
(ulong) srv_fatal_semaphore_wait_threshold);
2418
/* Flush stderr so that a database user gets the output
2419
to possible MySQL error file */
2423
os_thread_sleep(1000000);
2425
if (srv_shutdown_state < SRV_SHUTDOWN_CLEANUP) {
2430
srv_error_monitor_active = FALSE;
2432
/* We count the number of threads in os_thread_exit(). A created
2433
thread should always use that to exit and not use return() to exit. */
2435
os_thread_exit(NULL);
2437
OS_THREAD_DUMMY_RETURN;
2440
/**********************************************************************//**
2441
Check whether any background thread is active.
2442
@return FALSE if all are are suspended or have exited. */
2445
srv_is_any_background_thread_active(void)
2446
/*=====================================*/
2451
mutex_enter(&kernel_mutex);
2453
for (i = SRV_COM; i <= SRV_MASTER; ++i) {
2454
if (srv_n_threads_active[i] != 0) {
2460
mutex_exit(&kernel_mutex);
2465
/*******************************************************************//**
2466
Tells the InnoDB server that there has been activity in the database
2467
and wakes up the master thread if it is suspended (not sleeping). Used
2468
in the MySQL interface. Note that there is a small chance that the master
2469
thread stays suspended (we do not protect our operation with the
2470
srv_sys_t->mutex, for performance reasons). */
2473
srv_active_wake_master_thread(void)
2474
/*===============================*/
2476
srv_activity_count++;
2478
if (srv_n_threads_active[SRV_MASTER] == 0) {
2480
mutex_enter(&kernel_mutex);
2482
srv_release_threads(SRV_MASTER, 1);
2484
mutex_exit(&kernel_mutex);
2488
/*******************************************************************//**
2489
Tells the purge thread that there has been activity in the database
2490
and wakes up the purge thread if it is suspended (not sleeping). Note
2491
that there is a small chance that the purge thread stays suspended
2492
(we do not protect our operation with the kernel mutex, for
2493
performace reasons). */
2496
srv_wake_purge_thread_if_not_active(void)
2497
/*=====================================*/
2499
ut_ad(!mutex_own(&kernel_mutex));
2501
if (srv_n_purge_threads > 0
2502
&& srv_n_threads_active[SRV_WORKER] == 0) {
2504
mutex_enter(&kernel_mutex);
2506
srv_release_threads(SRV_WORKER, 1);
2508
mutex_exit(&kernel_mutex);
2512
/*******************************************************************//**
2513
Wakes up the master thread if it is suspended or being suspended. */
2516
srv_wake_master_thread(void)
2517
/*========================*/
2519
srv_activity_count++;
2521
mutex_enter(&kernel_mutex);
2523
srv_release_threads(SRV_MASTER, 1);
2525
mutex_exit(&kernel_mutex);
2528
/*******************************************************************//**
2529
Wakes up the purge thread if it's not already awake. */
2532
srv_wake_purge_thread(void)
2533
/*=======================*/
2535
ut_ad(!mutex_own(&kernel_mutex));
2537
if (srv_n_purge_threads > 0) {
2539
mutex_enter(&kernel_mutex);
2541
srv_release_threads(SRV_WORKER, 1);
2543
mutex_exit(&kernel_mutex);
2547
/**********************************************************************
2548
The master thread is tasked to ensure that flush of log file happens
2549
once every second in the background. This is to ensure that not more
2550
than one second of trxs are lost in case of crash when
2551
innodb_flush_logs_at_trx_commit != 1 */
2554
srv_sync_log_buffer_in_background(void)
2555
/*===================================*/
2557
time_t current_time = time(NULL);
2559
srv_main_thread_op_info = "flushing log";
2560
if (difftime(current_time, srv_last_log_flush_time) >= 1) {
2561
log_buffer_sync_in_background(TRUE);
2562
srv_last_log_flush_time = current_time;
2563
srv_log_writes_and_flush++;
2567
/********************************************************************//**
2568
Do a full purge, reconfigure the purge sub-system if a dynamic
2569
change is detected. */
2572
srv_master_do_purge(void)
2573
/*=====================*/
2575
ulint n_pages_purged;
2577
ut_ad(!mutex_own(&kernel_mutex));
2579
ut_a(srv_n_purge_threads == 0);
2582
/* Check for shutdown and change in purge config. */
2583
if (srv_fast_shutdown && srv_shutdown_state > 0) {
2584
/* Nothing to purge. */
2587
n_pages_purged = trx_purge(srv_purge_batch_size);
2590
srv_sync_log_buffer_in_background();
2592
} while (n_pages_purged > 0);
2595
/*********************************************************************//**
2596
The master thread controlling the server.
2597
@return a dummy parameter */
2602
void* arg __attribute__((unused)))
2603
/*!< in: a dummy parameter required by
2606
buf_pool_stat_t buf_stat;
2608
ulint old_activity_count;
2609
ulint n_pages_purged = 0;
2610
ulint n_bytes_merged;
2611
ulint n_pages_flushed;
2612
ulint n_bytes_archived;
2613
ulint n_tables_to_drop;
2616
ulint n_ios_very_old;
2618
ulint next_itr_time;
2621
#ifdef UNIV_DEBUG_THREAD_CREATION
2622
fprintf(stderr, "Master thread starts, id %lu\n",
2623
os_thread_pf(os_thread_get_curr_id()));
2626
#ifdef UNIV_PFS_THREAD
2627
pfs_register_thread(srv_master_thread_key);
2630
srv_main_thread_process_no = os_proc_get_number();
2631
srv_main_thread_id = os_thread_pf(os_thread_get_curr_id());
2633
srv_table_reserve_slot(SRV_MASTER);
2635
mutex_enter(&kernel_mutex);
2637
srv_n_threads_active[SRV_MASTER]++;
2639
mutex_exit(&kernel_mutex);
2642
/*****************************************************************/
2643
/* ---- When there is database activity by users, we cycle in this
2646
srv_main_thread_op_info = "reserving kernel mutex";
2648
buf_get_total_stat(&buf_stat);
2649
n_ios_very_old = log_sys->n_log_ios + buf_stat.n_pages_read
2650
+ buf_stat.n_pages_written;
2651
mutex_enter(&kernel_mutex);
2653
/* Store the user activity counter at the start of this loop */
2654
old_activity_count = srv_activity_count;
2656
mutex_exit(&kernel_mutex);
2658
if (srv_force_recovery >= SRV_FORCE_NO_BACKGROUND) {
2660
goto suspend_thread;
2663
/* ---- We run the following loop approximately once per second
2664
when there is database activity */
2666
srv_last_log_flush_time = time(NULL);
2668
/* Sleep for 1 second on entrying the for loop below the first time. */
2669
next_itr_time = ut_time_ms() + 1000;
2671
for (i = 0; i < 10; i++) {
2672
ulint cur_time = ut_time_ms();
2674
buf_get_total_stat(&buf_stat);
2676
n_ios_old = log_sys->n_log_ios + buf_stat.n_pages_read
2677
+ buf_stat.n_pages_written;
2679
srv_main_thread_op_info = "sleeping";
2680
srv_main_1_second_loops++;
2682
if (next_itr_time > cur_time) {
2684
/* Get sleep interval in micro seconds. We use
2685
ut_min() to avoid long sleep in case of
2687
os_thread_sleep(ut_min(1000000,
2688
(next_itr_time - cur_time)
2693
/* Each iteration should happen at 1 second interval. */
2694
next_itr_time = ut_time_ms() + 1000;
2696
/* ALTER TABLE in MySQL requires on Unix that the table handler
2697
can drop tables lazily after there no longer are SELECT
2700
srv_main_thread_op_info = "doing background drop tables";
2702
row_drop_tables_for_mysql_in_background();
2704
srv_main_thread_op_info = "";
2706
if (srv_fast_shutdown && srv_shutdown_state > 0) {
2708
goto background_loop;
2711
/* Flush logs if needed */
2712
srv_sync_log_buffer_in_background();
2714
srv_main_thread_op_info = "making checkpoint";
2717
/* If i/os during one second sleep were less than 5% of
2718
capacity, we assume that there is free disk i/o capacity
2719
available, and it makes sense to do an insert buffer merge. */
2721
buf_get_total_stat(&buf_stat);
2722
n_pend_ios = buf_get_n_pending_ios()
2723
+ log_sys->n_pending_writes;
2724
n_ios = log_sys->n_log_ios + buf_stat.n_pages_read
2725
+ buf_stat.n_pages_written;
2726
if (n_pend_ios < SRV_PEND_IO_THRESHOLD
2727
&& (n_ios - n_ios_old < SRV_RECENT_IO_ACTIVITY)) {
2728
srv_main_thread_op_info = "doing insert buffer merge";
2729
ibuf_contract_for_n_pages(FALSE, PCT_IO(5));
2731
/* Flush logs if needed */
2732
srv_sync_log_buffer_in_background();
2735
if (UNIV_UNLIKELY(buf_get_modified_ratio_pct()
2736
> srv_max_buf_pool_modified_pct)) {
2738
/* Try to keep the number of modified pages in the
2739
buffer pool under the limit wished by the user */
2741
srv_main_thread_op_info =
2742
"flushing buffer pool pages";
2743
n_pages_flushed = buf_flush_list(
2744
PCT_IO(100), IB_ULONGLONG_MAX);
2746
} else if (srv_adaptive_flushing) {
2748
/* Try to keep the rate of flushing of dirty
2749
pages such that redo log generation does not
2750
produce bursts of IO at checkpoint time. */
2751
ulint n_flush = buf_flush_get_desired_flush_rate();
2754
srv_main_thread_op_info =
2755
"flushing buffer pool pages";
2756
n_flush = ut_min(PCT_IO(100), n_flush);
2764
if (srv_activity_count == old_activity_count) {
2766
/* There is no user activity at the moment, go to
2767
the background loop */
2769
goto background_loop;
2773
/* ---- We perform the following code approximately once per
2774
10 seconds when there is database activity */
2776
#ifdef MEM_PERIODIC_CHECK
2777
/* Check magic numbers of every allocated mem block once in 10
2779
mem_validate_all_blocks();
2781
/* If i/os during the 10 second period were less than 200% of
2782
capacity, we assume that there is free disk i/o capacity
2783
available, and it makes sense to flush srv_io_capacity pages.
2785
Note that this is done regardless of the fraction of dirty
2786
pages relative to the max requested by the user. The one second
2787
loop above requests writes for that case. The writes done here
2788
are not required, and may be disabled. */
2790
buf_get_total_stat(&buf_stat);
2791
n_pend_ios = buf_get_n_pending_ios() + log_sys->n_pending_writes;
2792
n_ios = log_sys->n_log_ios + buf_stat.n_pages_read
2793
+ buf_stat.n_pages_written;
2795
srv_main_10_second_loops++;
2796
if (n_pend_ios < SRV_PEND_IO_THRESHOLD
2797
&& (n_ios - n_ios_very_old < SRV_PAST_IO_ACTIVITY)) {
2799
srv_main_thread_op_info = "flushing buffer pool pages";
2800
buf_flush_list(PCT_IO(100), IB_ULONGLONG_MAX);
2802
/* Flush logs if needed */
2803
srv_sync_log_buffer_in_background();
2806
/* We run a batch of insert buffer merge every 10 seconds,
2807
even if the server were active */
2809
srv_main_thread_op_info = "doing insert buffer merge";
2810
ibuf_contract_for_n_pages(FALSE, PCT_IO(5));
2812
/* Flush logs if needed */
2813
srv_sync_log_buffer_in_background();
2815
if (srv_n_purge_threads == 0) {
2816
srv_main_thread_op_info = "master purging";
2818
srv_master_do_purge();
2820
if (srv_fast_shutdown && srv_shutdown_state > 0) {
2822
goto background_loop;
2826
srv_main_thread_op_info = "flushing buffer pool pages";
2828
/* Flush a few oldest pages to make a new checkpoint younger */
2830
if (buf_get_modified_ratio_pct() > 70) {
2832
/* If there are lots of modified pages in the buffer pool
2833
(> 70 %), we assume we can afford reserving the disk(s) for
2834
the time it requires to flush 100 pages */
2836
n_pages_flushed = buf_flush_list(
2837
PCT_IO(100), IB_ULONGLONG_MAX);
2839
/* Otherwise, we only flush a small number of pages so that
2840
we do not unnecessarily use much disk i/o capacity from
2843
n_pages_flushed = buf_flush_list(
2844
PCT_IO(10), IB_ULONGLONG_MAX);
2847
srv_main_thread_op_info = "making checkpoint";
2849
/* Make a new checkpoint about once in 10 seconds */
2851
log_checkpoint(TRUE, FALSE);
2853
srv_main_thread_op_info = "reserving kernel mutex";
2855
mutex_enter(&kernel_mutex);
2857
/* ---- When there is database activity, we jump from here back to
2858
the start of loop */
2860
if (srv_activity_count != old_activity_count) {
2861
mutex_exit(&kernel_mutex);
2865
mutex_exit(&kernel_mutex);
2867
/* If the database is quiet, we enter the background loop */
2869
/*****************************************************************/
2871
/* ---- In this loop we run background operations when the server
2872
is quiet from user activity. Also in the case of a shutdown, we
2873
loop here, flushing the buffer pool to the data files. */
2875
/* The server has been quiet for a while: start running background
2877
srv_main_background_loops++;
2878
srv_main_thread_op_info = "doing background drop tables";
2880
n_tables_to_drop = row_drop_tables_for_mysql_in_background();
2882
if (n_tables_to_drop > 0) {
2883
/* Do not monopolize the CPU even if there are tables waiting
2884
in the background drop queue. (It is essentially a bug if
2885
MySQL tries to drop a table while there are still open handles
2886
to it and we had to put it to the background drop queue.) */
2888
os_thread_sleep(100000);
2891
if (srv_n_purge_threads == 0) {
2892
srv_main_thread_op_info = "master purging";
2894
srv_master_do_purge();
2897
srv_main_thread_op_info = "reserving kernel mutex";
2899
mutex_enter(&kernel_mutex);
2900
if (srv_activity_count != old_activity_count) {
2901
mutex_exit(&kernel_mutex);
2904
mutex_exit(&kernel_mutex);
2906
srv_main_thread_op_info = "doing insert buffer merge";
2908
if (srv_fast_shutdown && srv_shutdown_state > 0) {
2911
/* This should do an amount of IO similar to the number of
2912
dirty pages that will be flushed in the call to
2913
buf_flush_list below. Otherwise, the system favors
2914
clean pages over cleanup throughput. */
2915
n_bytes_merged = ibuf_contract_for_n_pages(FALSE,
2919
srv_main_thread_op_info = "reserving kernel mutex";
2921
mutex_enter(&kernel_mutex);
2922
if (srv_activity_count != old_activity_count) {
2923
mutex_exit(&kernel_mutex);
2926
mutex_exit(&kernel_mutex);
2929
srv_main_thread_op_info = "flushing buffer pool pages";
2930
srv_main_flush_loops++;
2931
if (srv_fast_shutdown < 2) {
2932
n_pages_flushed = buf_flush_list(
2933
PCT_IO(100), IB_ULONGLONG_MAX);
2935
/* In the fastest shutdown we do not flush the buffer pool
2936
to data files: we set n_pages_flushed to 0 artificially. */
2938
n_pages_flushed = 0;
2941
srv_main_thread_op_info = "reserving kernel mutex";
2943
mutex_enter(&kernel_mutex);
2944
if (srv_activity_count != old_activity_count) {
2945
mutex_exit(&kernel_mutex);
2948
mutex_exit(&kernel_mutex);
2950
srv_main_thread_op_info = "waiting for buffer pool flush to end";
2951
buf_flush_wait_batch_end(NULL, BUF_FLUSH_LIST);
2953
/* Flush logs if needed */
2954
srv_sync_log_buffer_in_background();
2956
srv_main_thread_op_info = "making checkpoint";
2958
log_checkpoint(TRUE, FALSE);
2960
if (buf_get_modified_ratio_pct() > srv_max_buf_pool_modified_pct) {
2962
/* Try to keep the number of modified pages in the
2963
buffer pool under the limit wished by the user */
2968
srv_main_thread_op_info = "reserving kernel mutex";
2970
mutex_enter(&kernel_mutex);
2971
if (srv_activity_count != old_activity_count) {
2972
mutex_exit(&kernel_mutex);
2975
mutex_exit(&kernel_mutex);
2977
srv_main_thread_op_info = "archiving log (if log archive is on)";
2979
log_archive_do(FALSE, &n_bytes_archived);
2981
n_bytes_archived = 0;
2983
/* Keep looping in the background loop if still work to do */
2985
if (srv_fast_shutdown && srv_shutdown_state > 0) {
2986
if (n_tables_to_drop + n_pages_flushed
2987
+ n_bytes_archived != 0) {
2989
/* If we are doing a fast shutdown (= the default)
2990
we do not do purge or insert buffer merge. But we
2991
flush the buffer pool completely to disk.
2992
In a 'very fast' shutdown we do not flush the buffer
2993
pool to data files: we have set n_pages_flushed to
2996
goto background_loop;
2998
} else if (n_tables_to_drop
2999
+ n_pages_purged + n_bytes_merged + n_pages_flushed
3000
+ n_bytes_archived != 0) {
3001
/* In a 'slow' shutdown we run purge and the insert buffer
3002
merge to completion */
3004
goto background_loop;
3007
/* There is no work for background operations either: suspend
3008
master thread to wait for more server activity */
3011
srv_main_thread_op_info = "suspending";
3013
mutex_enter(&kernel_mutex);
3015
if (row_get_background_drop_list_len_low() > 0) {
3016
mutex_exit(&kernel_mutex);
3021
event = srv_suspend_thread();
3023
mutex_exit(&kernel_mutex);
3025
/* DO NOT CHANGE THIS STRING. innobase_start_or_create_for_mysql()
3026
waits for database activity to die down when converting < 4.1.x
3027
databases, and relies on this string being exactly as it is. InnoDB
3028
manual also mentions this string in several places. */
3029
srv_main_thread_op_info = "waiting for server activity";
3031
os_event_wait(event);
3033
if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
3034
/* This is only extra safety, the thread should exit
3035
already when the event wait ends */
3037
os_thread_exit(NULL);
3041
/* When there is user activity, InnoDB will set the event and the
3042
main thread goes back to loop. */
3047
#if (!defined(__SUNPRO_C) && !defined(__SUNPRO_CC))
3048
OS_THREAD_DUMMY_RETURN; /* Not reached, avoid compiler warning */
3052
/*********************************************************************//**
3053
Asynchronous purge thread.
3054
@return a dummy parameter */
3059
void* arg __attribute__((unused))) /*!< in: a dummy parameter
3060
required by os_thread_create */
3063
ulint slot_no = ULINT_UNDEFINED;
3064
ulint n_total_purged = ULINT_UNDEFINED;
3066
ut_a(srv_n_purge_threads == 1);
3068
#ifdef UNIV_DEBUG_THREAD_CREATION
3069
fprintf(stderr, "InnoDB: Purge thread running, id %lu\n",
3070
os_thread_pf(os_thread_get_curr_id()));
3071
#endif /* UNIV_DEBUG_THREAD_CREATION */
3073
mutex_enter(&kernel_mutex);
3075
slot_no = srv_table_reserve_slot(SRV_WORKER);
3077
slot = srv_table_get_nth_slot(slot_no);
3079
++srv_n_threads_active[SRV_WORKER];
3081
mutex_exit(&kernel_mutex);
3083
while (srv_shutdown_state != SRV_SHUTDOWN_EXIT_THREADS) {
3085
ulint n_pages_purged;
3087
/* If there are very few records to purge or the last
3088
purge didn't purge any records then wait for activity.
3089
We peek at the history len without holding any mutex
3090
because in the worst case we will end up waiting for
3091
the next purge event. */
3092
if (trx_sys->rseg_history_len < srv_purge_batch_size
3093
|| n_total_purged == 0) {
3097
mutex_enter(&kernel_mutex);
3099
event = srv_suspend_thread();
3101
mutex_exit(&kernel_mutex);
3103
os_event_wait(event);
3106
/* Check for shutdown and whether we should do purge at all. */
3107
if (srv_force_recovery >= SRV_FORCE_NO_BACKGROUND
3108
|| srv_shutdown_state != 0
3109
|| srv_fast_shutdown) {
3116
/* Purge until there are no more records to purge and there is
3117
no change in configuration or server state. */
3119
n_pages_purged = trx_purge(srv_purge_batch_size);
3121
n_total_purged += n_pages_purged;
3123
} while (n_pages_purged > 0 && !srv_fast_shutdown);
3125
srv_sync_log_buffer_in_background();
3128
mutex_enter(&kernel_mutex);
3130
ut_ad(srv_table_get_nth_slot(slot_no) == slot);
3132
/* Decrement the active count. */
3133
srv_suspend_thread();
3135
slot->in_use = FALSE;
3137
/* Free the thread local memory. */
3138
thr_local_free(os_thread_get_curr_id());
3140
mutex_exit(&kernel_mutex);
3142
#ifdef UNIV_DEBUG_THREAD_CREATION
3143
fprintf(stderr, "InnoDB: Purge thread exiting, id %lu\n",
3144
os_thread_pf(os_thread_get_curr_id()));
3145
#endif /* UNIV_DEBUG_THREAD_CREATION */
3147
/* We count the number of threads in os_thread_exit(). A created
3148
thread should always use that to exit and not use return() to exit. */
3149
os_thread_exit(NULL);
3151
OS_THREAD_DUMMY_RETURN; /* Not reached, avoid compiler warning */
3154
/**********************************************************************//**
3155
Enqueues a task to server task queue and releases a worker thread, if there
3156
is a suspended one. */
3159
srv_que_task_enqueue_low(
3160
/*=====================*/
3161
que_thr_t* thr) /*!< in: query thread */
3165
mutex_enter(&kernel_mutex);
3167
UT_LIST_ADD_LAST(queue, srv_sys->tasks, thr);
3169
srv_release_threads(SRV_WORKER, 1);
3171
mutex_exit(&kernel_mutex);