srv0srv.c, log0log.c, srv0srv.h, ha_innodb.cc, ha_innodb.h:

  Implement innobase_very_fast_shutdown and innobase_start_trx_and_assign_read_view(); these were requested by Guilhem
parent 6809db3d
...@@ -100,7 +100,10 @@ extern ulint srv_max_n_threads; ...@@ -100,7 +100,10 @@ extern ulint srv_max_n_threads;
extern lint srv_conc_n_threads; extern lint srv_conc_n_threads;
extern ibool srv_fast_shutdown; extern ibool srv_fast_shutdown;
extern ibool srv_very_fast_shutdown; /* if this TRUE, do not flush the
buffer pool to data files at the
shutdown; we effectively 'crash'
InnoDB */
extern ibool srv_innodb_status; extern ibool srv_innodb_status;
extern ibool srv_use_doublewrite_buf; extern ibool srv_use_doublewrite_buf;
......
...@@ -3048,13 +3048,25 @@ loop: ...@@ -3048,13 +3048,25 @@ loop:
#ifdef UNIV_LOG_ARCHIVE #ifdef UNIV_LOG_ARCHIVE
log_archive_all(); log_archive_all();
#endif /* UNIV_LOG_ARCHIVE */ #endif /* UNIV_LOG_ARCHIVE */
if (!srv_very_fast_shutdown) {
/* In a 'very fast' shutdown we do not flush the buffer pool:
it is essentially a 'crash' of the InnoDB server. */
log_make_checkpoint_at(ut_dulint_max, TRUE); log_make_checkpoint_at(ut_dulint_max, TRUE);
} else {
/* Make sure that the log is all flushed to disk, so that
we can recover all committed transactions in a crash
recovery */
log_buffer_flush_to_disk();
}
mutex_enter(&(log_sys->mutex)); mutex_enter(&(log_sys->mutex));
lsn = log_sys->lsn; lsn = log_sys->lsn;
if (ut_dulint_cmp(lsn, log_sys->last_checkpoint_lsn) != 0 if ((ut_dulint_cmp(lsn, log_sys->last_checkpoint_lsn) != 0
&& !srv_very_fast_shutdown)
#ifdef UNIV_LOG_ARCHIVE #ifdef UNIV_LOG_ARCHIVE
|| (srv_log_archive_on || (srv_log_archive_on
&& ut_dulint_cmp(lsn, && ut_dulint_cmp(lsn,
...@@ -3098,11 +3110,12 @@ loop: ...@@ -3098,11 +3110,12 @@ loop:
fil_flush_file_spaces(FIL_TABLESPACE); fil_flush_file_spaces(FIL_TABLESPACE);
fil_flush_file_spaces(FIL_LOG); fil_flush_file_spaces(FIL_LOG);
/* The next fil_write_... will pass the buffer pool: therefore /* The call fil_write_flushed_lsn_to_data_files() will pass the buffer
it is essential that the buffer pool has been completely flushed pool: therefore it is essential that the buffer pool has been
to disk! */ completely flushed to disk! (We do not call fil_write... if the
'very fast' shutdown is enabled.) */
if (!buf_all_freed()) { if (!srv_very_fast_shutdown && !buf_all_freed()) {
goto loop; goto loop;
} }
...@@ -3125,7 +3138,7 @@ loop: ...@@ -3125,7 +3138,7 @@ loop:
/* Make some checks that the server really is quiet */ /* Make some checks that the server really is quiet */
ut_a(srv_n_threads_active[SRV_MASTER] == 0); ut_a(srv_n_threads_active[SRV_MASTER] == 0);
ut_a(buf_all_freed()); ut_a(srv_very_fast_shutdown || buf_all_freed());
ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn)); ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn));
if (ut_dulint_cmp(lsn, srv_start_lsn) < 0) { if (ut_dulint_cmp(lsn, srv_start_lsn) < 0) {
...@@ -3140,7 +3153,15 @@ loop: ...@@ -3140,7 +3153,15 @@ loop:
srv_shutdown_lsn = lsn; srv_shutdown_lsn = lsn;
if (!srv_very_fast_shutdown) {
/* In a 'very fast' shutdown we do not flush the buffer pool:
it is essentially a 'crash' of the InnoDB server. Then we must
not write the lsn stamps to the data files, since at a
startup InnoDB deduces from the stamps if the previous
shutdown was clean. */
fil_write_flushed_lsn_to_data_files(lsn, arch_log_no); fil_write_flushed_lsn_to_data_files(lsn, arch_log_no);
}
fil_flush_file_spaces(FIL_TABLESPACE); fil_flush_file_spaces(FIL_TABLESPACE);
...@@ -3148,7 +3169,7 @@ loop: ...@@ -3148,7 +3169,7 @@ loop:
/* Make some checks that the server really is quiet */ /* Make some checks that the server really is quiet */
ut_a(srv_n_threads_active[SRV_MASTER] == 0); ut_a(srv_n_threads_active[SRV_MASTER] == 0);
ut_a(buf_all_freed()); ut_a(srv_very_fast_shutdown || buf_all_freed());
ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn)); ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn));
} }
......
...@@ -249,6 +249,10 @@ merge to completion before shutdown */ ...@@ -249,6 +249,10 @@ merge to completion before shutdown */
ibool srv_fast_shutdown = FALSE; ibool srv_fast_shutdown = FALSE;
ibool srv_very_fast_shutdown = FALSE; /* if this TRUE, do not flush the
buffer pool to data files at the
shutdown; we effectively 'crash'
InnoDB */
/* Generate a innodb_status.<pid> file */ /* Generate a innodb_status.<pid> file */
ibool srv_innodb_status = FALSE; ibool srv_innodb_status = FALSE;
...@@ -2178,7 +2182,8 @@ loop: ...@@ -2178,7 +2182,8 @@ loop:
/*****************************************************************/ /*****************************************************************/
background_loop: background_loop:
/* ---- In this loop we run background operations when the server /* ---- In this loop we run background operations when the server
is quiet from user activity */ is quiet from user activity. Also in the case of a shutdown, we
loop here, flushing the buffer pool to the data files. */
/* The server has been quiet for a while: start running background /* The server has been quiet for a while: start running background
operations */ operations */
...@@ -2251,7 +2256,16 @@ background_loop: ...@@ -2251,7 +2256,16 @@ background_loop:
flush_loop: flush_loop:
srv_main_thread_op_info = "flushing buffer pool pages"; srv_main_thread_op_info = "flushing buffer pool pages";
n_pages_flushed = buf_flush_batch(BUF_FLUSH_LIST, 100, ut_dulint_max);
if (!srv_very_fast_shutdown) {
n_pages_flushed =
buf_flush_batch(BUF_FLUSH_LIST, 100, ut_dulint_max);
} else {
/* In a 'very fast' shutdown we do not flush the buffer pool
to data files: we set n_pages_flushed to 0 artificially. */
n_pages_flushed = 0;
}
srv_main_thread_op_info = "reserving kernel mutex"; srv_main_thread_op_info = "reserving kernel mutex";
...@@ -2304,7 +2318,10 @@ flush_loop: ...@@ -2304,7 +2318,10 @@ flush_loop:
/* If we are doing a fast shutdown (= the default) /* If we are doing a fast shutdown (= the default)
we do not do purge or insert buffer merge. But we we do not do purge or insert buffer merge. But we
flush the buffer pool completely to disk. */ flush the buffer pool completely to disk.
In a 'very fast' shutdown we do not flush the buffer
pool to data files: we have set n_pages_flushed to
0 artificially. */
goto background_loop; goto background_loop;
} }
......
...@@ -20,6 +20,9 @@ NOTE: You can only use noninlined InnoDB functions in this file, because we ...@@ -20,6 +20,9 @@ NOTE: You can only use noninlined InnoDB functions in this file, because we
have disables the InnoDB inlining in this file. */ have disables the InnoDB inlining in this file. */
/* TODO list for the InnoDB handler in 4.1: /* TODO list for the InnoDB handler in 4.1:
- Remove the flag innodb_active_trans from thd and replace it with a
function call innodb_active_trans(thd), which looks at the InnoDB
trx struct state field
- Find out what kind of problems the OS X case-insensitivity causes to - Find out what kind of problems the OS X case-insensitivity causes to
table and database names; should we 'normalize' the names like we do table and database names; should we 'normalize' the names like we do
in Windows? in Windows?
...@@ -114,6 +117,9 @@ uint innobase_flush_log_at_trx_commit = 1; ...@@ -114,6 +117,9 @@ uint innobase_flush_log_at_trx_commit = 1;
my_bool innobase_log_archive = FALSE;/* unused */ my_bool innobase_log_archive = FALSE;/* unused */
my_bool innobase_use_native_aio = FALSE; my_bool innobase_use_native_aio = FALSE;
my_bool innobase_fast_shutdown = TRUE; my_bool innobase_fast_shutdown = TRUE;
my_bool innobase_very_fast_shutdown = FALSE; /* this can be set to
1 just prior calling
innobase_end() */
my_bool innobase_file_per_table = FALSE; my_bool innobase_file_per_table = FALSE;
my_bool innobase_locks_unsafe_for_binlog = FALSE; my_bool innobase_locks_unsafe_for_binlog = FALSE;
my_bool innobase_create_status_file = FALSE; my_bool innobase_create_status_file = FALSE;
...@@ -799,6 +805,10 @@ ha_innobase::init_table_handle_for_HANDLER(void) ...@@ -799,6 +805,10 @@ ha_innobase::init_table_handle_for_HANDLER(void)
trx_assign_read_view(prebuilt->trx); trx_assign_read_view(prebuilt->trx);
/* Set the MySQL flag to mark that there is an active transaction */
current_thd->transaction.all.innodb_active_trans = 1;
/* We did the necessary inits in this function, no need to repeat them /* We did the necessary inits in this function, no need to repeat them
in row_search_for_mysql */ in row_search_for_mysql */
...@@ -1059,6 +1069,15 @@ innobase_end(void) ...@@ -1059,6 +1069,15 @@ innobase_end(void)
#endif #endif
if (innodb_inited) if (innodb_inited)
{ {
if (innobase_very_fast_shutdown) {
srv_very_fast_shutdown = TRUE;
fprintf(stderr,
"InnoDB: MySQL has requested a very fast shutdown without flushing\n"
"InnoDB: the InnoDB buffer pool to data files. At the next mysqld startup\n"
"InnoDB: InnoDB will do a crash recovery!\n");
}
innodb_inited= 0; innodb_inited= 0;
if (innobase_shutdown_for_mysql() != DB_SUCCESS) if (innobase_shutdown_for_mysql() != DB_SUCCESS)
err= 1; err= 1;
...@@ -1115,6 +1134,48 @@ innobase_commit_low( ...@@ -1115,6 +1134,48 @@ innobase_commit_low(
trx_commit_for_mysql(trx); trx_commit_for_mysql(trx);
} }
/*********************************************************************
Creates an InnoDB transaction struct for the thd if it does not yet have one.
Starts a new InnoDB transaction if a transaction is not yet started. And
assigns a new snapshot for a consistent read if the transaction does not yet
have one. */
int
innobase_start_trx_and_assign_read_view(
/*====================================*/
/* out: 0 */
THD* thd) /* in: MySQL thread handle of the user for whom
the transaction should be committed */
{
trx_t* trx;
DBUG_ENTER("innobase_start_trx_and_assign_read_view");
/* Create a new trx struct for thd, if it does not yet have one */
trx = check_trx_exists(thd);
/* This is just to play safe: release a possible FIFO ticket and
search latch. Since we will reserve the kernel mutex, we have to
release the search system latch first to obey the latching order. */
innobase_release_stat_resources(trx);
/* If the transaction is not started yet, start it */
trx_start_if_not_started_noninline(trx);
/* Assign a read view if the transaction does not have it yet */
trx_assign_read_view(trx);
/* Set the MySQL flag to mark that there is an active transaction */
current_thd->transaction.all.innodb_active_trans = 1;
DBUG_RETURN(0);
}
/********************************************************************* /*********************************************************************
Commits a transaction in an InnoDB database or marks an SQL statement Commits a transaction in an InnoDB database or marks an SQL statement
ended. */ ended. */
...@@ -1146,8 +1207,10 @@ innobase_commit( ...@@ -1146,8 +1207,10 @@ innobase_commit(
1. ::external_lock(), 1. ::external_lock(),
2. ::start_stmt(), 2. ::start_stmt(),
3. innobase_query_caching_of_table_permitted(), and 3. innobase_query_caching_of_table_permitted(),
4. innobase_savepoint(), 4. innobase_savepoint(),
5. ::init_table_handle_for_HANDLER(),
6. innobase_start_trx_and_assign_read_view()
and it is only set to 0 in a commit or a rollback. If it is 0 we know and it is only set to 0 in a commit or a rollback. If it is 0 we know
there cannot be resources to be freed and we could return immediately. there cannot be resources to be freed and we could return immediately.
......
...@@ -191,6 +191,11 @@ extern my_bool innobase_log_archive, ...@@ -191,6 +191,11 @@ extern my_bool innobase_log_archive,
innobase_use_native_aio, innobase_fast_shutdown, innobase_use_native_aio, innobase_fast_shutdown,
innobase_file_per_table, innobase_locks_unsafe_for_binlog, innobase_file_per_table, innobase_locks_unsafe_for_binlog,
innobase_create_status_file; innobase_create_status_file;
extern my_bool innobase_very_fast_shutdown; /* set this to 1 just before
calling innobase_end() if you want
InnoDB to shut down without
flushing the buffer pool: this
is equivalent to a 'crash' */
extern "C" { extern "C" {
extern ulong srv_max_buf_pool_modified_pct; extern ulong srv_max_buf_pool_modified_pct;
extern ulong srv_auto_extend_increment; extern ulong srv_auto_extend_increment;
...@@ -231,3 +236,4 @@ void innobase_release_temporary_latches(void* innobase_tid); ...@@ -231,3 +236,4 @@ void innobase_release_temporary_latches(void* innobase_tid);
void innobase_store_binlog_offset_and_flush_log(char *binlog_name,longlong offset); void innobase_store_binlog_offset_and_flush_log(char *binlog_name,longlong offset);
int innobase_start_trx_and_assign_read_view(THD* thd);
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment