This bug was originally filed and fixed as Bug#12612184. The original
fix was buggy, and it was patched by Bug#12704861. Also that patch was
buggy (potentially breaking crash recovery), and both fixes were
reverted.
This fix was not ported to the built-in InnoDB of MySQL 5.1, because
the function signatures of many core functions are different from
InnoDB Plugin and later versions. The block allocation routines and
their callers would have to changed so that they handle block
descriptors instead of page frames.
When a record is updated so that its size grows, non-updated columns
can be selected for external (off-page) storage. The bug is that the
initially inserted updated record contains an all-zero BLOB pointer to
the field that was not updated. Only after the BLOB pages have been
allocated and written, the valid pointer can be written to the record.
Between the release of the page latch in mtr_commit(mtr) after
btr_cur_pessimistic_update() and the re-latching of the page in
btr_pcur_restore_position(), other threads can see the invalid BLOB
pointer consisting of 20 zero bytes. Moreover, if the system crashes
at this point, the situation could persist after crash recovery, and
the contents of the non-updated column would be permanently lost.
The problem is amplified by the ROW_FORMAT=DYNAMIC and
ROW_FORMAT=COMPRESSED that were introduced in
innodb_file_format=barracuda in InnoDB Plugin, but the bug does exist
in all InnoDB versions.
The fix is as follows. After a pessimistic B-tree operation that needs
to write out off-page columns, allocate the pages for these columns in
the mini-transaction that performed the B-tree operation (btr_mtr),
but write the pages in a separate mini-transaction (blob_mtr). Do
mtr_commit(blob_mtr) before mtr_commit(btr_mtr). A quirk: Do not reuse
pages that were previously freed in btr_mtr. Only write the off-page
columns to 'fresh' pages.
In this way, crash recovery will see redo log entries for blob_mtr
before any redo log entry for btr_mtr. It will apply the BLOB page
writes to pages that were marked free at that point. If crash recovery
fails to see all of the btr_mtr redo log, there will be some
unreachable BLOB data in free pages, but the B-tree will be in a
consistent state.
btr_page_alloc_low(): Renamed from btr_page_alloc(). Add the parameter
init_mtr. Return an allocated block, or NULL. If init_mtr!=mtr but
the page was already X-latched in mtr, do not initialize the page.
btr_page_alloc(): Wrapper for btr_page_alloc_for_ibuf() and
btr_page_alloc_low().
btr_page_free(): Add a debug assertion that the page was a B-tree page.
btr_lift_page_up(): Return the father block.
btr_compress(), btr_cur_compress_if_useful(): Add the parameter ibool
adjust, for adjusting the cursor position.
btr_cur_pessimistic_update(): Preserve the cursor position when
big_rec will be written and the new flag BTR_KEEP_POS_FLAG is defined.
Remove a duplicate rec_get_offsets() call. Keep the X-latch on
index->lock when big_rec is needed.
btr_store_big_rec_extern_fields(): Replace update_inplace with
an operation code, and local_mtr with btr_mtr. When not doing a
fresh insert and btr_mtr has freed pages, put aside any pages that
were previously X-latched in btr_mtr, and free the pages after
writing out all data. The data must be written to 'fresh' pages,
because btr_mtr will be committed and written to the redo log after
the BLOB writes have been written to the redo log.
btr_blob_op_is_update(): Check if an operation passed to
btr_store_big_rec_extern_fields() is an update or insert-by-update.
fseg_alloc_free_page_low(), fsp_alloc_free_page(),
fseg_alloc_free_extent(), fseg_alloc_free_page_general(): Add the
parameter init_mtr. Return an allocated block, or NULL. If
init_mtr!=mtr but the page was already X-latched in mtr, do not
initialize the page.
xdes_get_descriptor_with_space_hdr(): Assert that the file space
header is being X-latched.
fsp_alloc_from_free_frag(): Refactored from fsp_alloc_free_page().
fsp_page_create(): New function, for allocating, X-latching and
potentially initializing a page. If init_mtr!=mtr but the page was
already X-latched in mtr, do not initialize the page.
fsp_free_page(): Add ut_ad(0) to the error outcomes.
fsp_free_page(), fseg_free_page_low(): Increment mtr->n_freed_pages.
fsp_alloc_seg_inode_page(), fseg_create_general(): Assert that the
page was not previously X-latched in the mini-transaction. A file
segment or inode page should never be allocated in the middle of an
mini-transaction that frees pages, such as btr_cur_pessimistic_delete().
fseg_alloc_free_page_low(): If the hinted page was allocated, skip the
check if the tablespace should be extended. Return NULL instead of
FIL_NULL on failure. Remove the flag frag_page_allocated. Instead,
return directly, because the page would already have been initialized.
fseg_find_free_frag_page_slot() would return ULINT_UNDEFINED on error,
not FIL_NULL. Correct a bogus assertion.
fseg_alloc_free_page(): Redefine as a wrapper macro around
fseg_alloc_free_page_general().
buf_block_buf_fix_inc(): Move the definition from the buf0buf.ic to
buf0buf.h, so that it can be called from other modules.
mtr_t: Add n_freed_pages (number of pages that have been freed).
page_rec_get_nth_const(), page_rec_get_nth(): The inverse function of
page_rec_get_n_recs_before(), get the nth record of the record
list. This is faster than iterating the linked list. Refactored from
page_get_middle_rec().
trx_undo_rec_copy(): Add a debug assertion for the length.
trx_undo_add_page(): Return a block descriptor or NULL instead of a
page number or FIL_NULL.
trx_undo_report_row_operation(): Add debug assertions.
trx_sys_create_doublewrite_buf(): Assert that each page was not
previously X-latched.
page_cur_insert_rec_zip_reorg(): Make use of page_rec_get_nth().
row_ins_clust_index_entry_by_modify(): Pass BTR_KEEP_POS_FLAG, so that
the repositioning of the cursor can be avoided.
row_ins_index_entry_low(): Add DEBUG_SYNC points before and after
writing off-page columns. If inserting by updating a delete-marked
record, do not reposition the cursor or commit the mini-transaction
before writing the off-page columns.
row_build(): Tighten a debug assertion about null BLOB pointers.
row_upd_clust_rec(): Add DEBUG_SYNC points before and after writing
off-page columns. Do not reposition the cursor or commit the
mini-transaction before writing the off-page columns.
rb:939 approved by Jimmy Yang
Problem: Statements that write to tables with auto_increment columns
based on the selection from another table, may lead to master
and slave going out of sync, as the order in which the rows
are retrieved from the table may differ on master and slave.
Solution: We mark writing to a table with auto_increment table
based on the rows selected from another table as unsafe. This
will cause the execution of such statements to throw a warning
and forces the statement to be logged in ROW if the logging
format is mixed.
Changes:
1. All the statements that writes to a table with auto_increment
column(s) based on the rows fetched from another table, will now
be unsafe.
2. CREATE TABLE with SELECT will now be unsafe.
Problem: Statements that write to tables with auto_increment columns
based on the selection from another table, may lead to master
and slave going out of sync, as the order in which the rows
are retrived from the table may differ on master and slave.
Solution: We mark writing to a table with auto_increment table
as unsafe. This will cause the execution of such statements to
throw a warning and forces the statement to be logged in ROW if
the logging format is mixed.
Changes:
1. All the statements that writes to a table with auto_increment
column(s) based on the rows fetched from another table, will now
be unsafe.
2. CREATE TABLE with SELECT will now be unsafe.
The actual Bug#11754376 does not exist in MySQL 5.5 because at startup
we drop entries for temporary tables from InnoDB dictionary cache (only
if ROW_FORMAT is not REDUNDANT). But nevertheless the bug in
normalize_table_name_low() is present so we fix it.
GRACEFUL SHUTDOWN
During startup mysql picks up .frm files from the tmpdir directory and
tries to drop those tables in the storage engine.
The problem is that when tmpdir ends in / then ha_innobase::delete_table()
is passed a string like "/var/tmp//#sql123", then it wrongly normalizes it
to "/#sql123" and calls row_drop_table_for_mysql() which of course fails
to delete the table entry from the InnoDB dictionary cache.
ha_innobase::delete_table() returns an error but nevertheless mysql wipes
away the .frm file and the entry in the InnoDB dictionary cache remains
orphaned with no easy way to remove it.
The "no easy" way to remove it is to create a similar temporary table again,
copy its .frm file to tmpdir under "#sql123.frm" and restart mysqld with
tmpdir=/var/tmp (no trailing slash) - this way mysql will pick the .frm file
after restart and will try to issue drop table for "/var/tmp/#sql123"
(notice do double slash), ha_innobase::delete_table() will normalize it to
"tmp/#sql123" and row_drop_table_for_mysql() will successfully remove the
table entry from the dictionary cache.
The solution is to fix normalize_table_name_low() to normalize things like
"/var/tmp//table" correctly to "tmp/table".
This patch also adds a test function which invokes
normalize_table_name_low() with various inputs to make sure it works
correctly and a mtr test that calls this test function.
Reviewed by: Marko (http://bur03.no.oracle.com/rb/r/929/)
rpl_heartbeat_basic test fails sporadically on pushbuild because did
not received all heartbeats from slave in circular replication.
MASTER_HEARTBEAT_PERIOD had the default value (slave_net_timeout/2) so
wait on "Heartbeat event received on master", that only waits for 1
minute, sometimes timeout before heartbeat arrives. Fixed setting a
smaller period value.
MEMORY LEAK.
Background:
- There are caches for stored functions and stored procedures (SP-cache);
- There is no similar cache for events;
- Triggers are cached together with TABLE objects;
- Those SP-caches are per-session (i.e. specific to each session);
- A stored routine is represented by a sp_head-instance internally;
- SP-cache basically contains sp_head-objects of stored routines, which
have been executed in a session;
- sp_head-object is added into the SP-cache before the corresponding
stored routine is executed;
- SP-cache is flushed in the end of the session.
The problem was that SP-cache might grow without any limit. Although this
was not a pure memory leak (the SP-cache is flushed when session is closed),
this is still a problem, because the user might take much memory by
executing many stored routines.
The patch fixes this problem in the least-intrusive way. A soft limit
(similar to the size of table definition cache) is introduced. To represent
such limit the new runtime configuration parameter 'stored_program_cache'
is introduced. The value of this parameter is stored in the new global
variable stored_program_cache_size that used to control the size of SP-cache
to overflow.
The parameter 'stored_program_cache' limits number of cached routines for
each thread. It has the following min/default/max values given from support:
min = 256, default = 256, max = 512 * 1024.
Also it should be noted that this parameter limits the size of
each cache (for stored procedures and for stored functions) separately.
The SP-cache size is checked after top-level statement is parsed.
If SP-cache size exceeds the limit specified by parameter
'stored_program_cache' then SP-cache is flushed and memory allocated for
cache objects is freed. Such approach allows to flush cache safely
when there are dependencies among stored routines.
Problem : The basic problem is the way the thread sleeps in mysql-5.5 and also in mysql-5.1
when we execute a stop slave on windows platform.
On windows platform if the stop slave is executed after the master dies, we have
this long wait before the stop slave return a value. This is because there is a
sleep of the thread. The sleep is uninterruptable in the two above version,
which was fixed by Davi patch for the BUG#11765860 for mysql-trunk. Backporting
his patch for mysql-5.5 fixes the problem.
Solution : A new pair of mutex and condition variable is introduced to synchronize thread
sleep and finalization. A new mutex is required because the slave threads are
terminated while holding the slave thread locks (run_lock), which can not be
relinquished during termination as this would affect the lock order.
If an error message contains '\' backslash it is displayed correctly
through show-slave-status or
query_get_value(SHOW SLAVE STATUS, Last_IO_Error, 1);. But when
SELECT REPLACE(...) is applied backslash is escaped resulting in a
different test output.
Disabled backslash escape on show_slave_status.inc and replaced '\' for
'/' using replace_regex function in order to achieve the same test
output when different path separators are used.
The server crashes when receiving a COM_BINLOG_DUMP command with a position of 0 or
larger than the file size.
The execution proceeds to an error block having the last read replication coordinates
pointer be NULL and its dereferencing crashed the server.
Fixed with making "public" previously used only for heartbeat coordinates.
and cryptic error 1126 message
The problem was that dlopen() related code was using just a subset
of the path normalization routines used in other places.
Fixed the expansion of the pre-dlopen() behavior for plugins and UDFs
to use a platform-dependent consistent encoding of the paths.
Fixed the error dlopen() error handling to take the correct error message
and strip off the trailing newline character(s).
Fixed tests to do a platform independent replace of directories and to
account for the traling slash.
Test extra/rpl_tests/rpl_extra_col_master.test (used by
rpl_extra_col_master_*) ends with the active connection pointing to the
slave. Thus, the two last tests never succeed in changing the binlog
format of the master away from 'row'. With correct active connection
(master) tests fail for binlog 'statement' and 'mixed' formats.
Tests rpl_extra_col_master_* only run when binary log format is
row. Statement and mix replication do not make sense in this
tests since it will try to execute statements on columns that do
not exist. This fix is basically a backport from mysql-5.5, see
changes done as part of BUG 39934.
If we meet DB_TOO_MANY_CONCURRENT_TRXS during the execution tab_create_graph from row_create_table_for_mysql(), .ibd file for the table should be created already but was not deleted for the error handling.
rb:875 approved by Jimmy Yang
If we meet DB_TOO_MANY_CONCURRENT_TRXS during the execution tab_create_graph from row_create_table_for_mysql(), .ibd file for the table should be created already but was not deleted for the error handling.
rb:875 approved by Jimmy Yang
CREATE TABLE bug13510739 (c INTEGER NOT NULL, PRIMARY KEY (c)) ENGINE=INNODB;
INSERT INTO bug13510739 VALUES (1), (2), (3), (4);
DELETE FROM bug13510739 WHERE c=2;
HANDLER bug13510739 OPEN;
HANDLER bug13510739 READ `primary` = (2);
HANDLER bug13510739 READ `primary` NEXT; <-- crash
The bug is that in the particular testcase row_search_for_mysql() picked up
a delete-marked record and quit, leaving the cursor non-positioned state and
on the subsequent 'get next' call the code crashed because of the
non-positioned cursor.
In row0sel.cc (line numbers from mysql-trunk):
4653 if (rec_get_deleted_flag(rec, comp)) {
...
4679 if (index == clust_index && unique_search) {
4680
4681 err = DB_RECORD_NOT_FOUND;
4682
4683 goto normal_return;
4684 }
it quit from here, not storing the cursor position.
In contrast, if the record=2 is not found at all (e.g. sleep(1) after DELETE
to let the purge wipe it away completely) then 'get = 2' does find record=3
and quits from here:
4366 if (0 != cmp_dtuple_rec(search_tuple, rec, offsets)) {
...
4394 btr_pcur_store_position(pcur, &mtr);
4395
4396 err = DB_RECORD_NOT_FOUND;
4397 #if 0
4398 ut_print_name(stderr, trx, FALSE, index->name);
4399 fputs(" record not found 3\n", stderr);
4400 #endif
4401
4402 goto normal_return;
Another fix could be to extend the condition on line 4366 to hold only if
seach_tuple matches rec AND if rec is not delete marked.
Notice that in the above test case if we wait about 1 second somewhere after
DELETE and before 'get = 2', then the testcase does not crash and returns 4
instead. Not sure if this is the correct behavior, but this bugfix removes
the crash and makes the code return what it also returns in the non-crashing
case (if rec=2 is not found during 'get = 2', e.g. we have sleep(1) there).
Approved by: Marko (http://bur03.no.oracle.com/rb/r/863/)
There was memory leak when running some tests on PB2.
The reason of the failure is an early return from change_master()
that was supposed to deallocate a dyn-array.
Actually the same bug58915 was fixed in trunk with relocating the dyn-array
destruction into THD::cleanup_after_query() which can't be bypassed.
The current patch backports magne.mahre@oracle.com-20110203101306-q8auashb3d7icxho
and adds two optimizations: were done: the static buffer for the dyn-array to base on,
and the array initialization is called precisely when it's necessary rather than
per each CHANGE-MASTER as before.
The counter handler_read_key (SSV::ha_read_key_count) is incremented
incorrectly.
The mysql server maintains a per thread system_status_var (SSV)
object. This object contains among other things the counter
SSV::ha_read_key_count. The purpose of this counter is to measure the
number of requests to read a row based on a key (or the number of
index lookups).
This counter was wrongly incremented in the
ha_innobase::innobase_get_index(). The fix removes
this increment statement (for both innodb and innodb_plugin).
The various callers of the innobase_get_index() was checked to
determine if anybody must increment this counter (if they first call
innobase_get_index() and then perform an index lookup). It was found
that no caller of innobase_get_index() needs to worry about the
SSV::ha_read_key_count counter.
WITH LARGE BUFFER POOL
(Note: this a backport of revno:3472 from mysql-trunk)
rb://845
approved by: Marko
When dropping a table (with an .ibd file i.e.: with
innodb_file_per_table set) we scan entire LRU to invalidate pages from
that table. This can be painful in case of large buffer pools as we hold
the buf_pool->mutex for the scan. Note that gravity of the problem does
not depend on the size of the table. Even with an empty table but a
large and filled up buffer pool we'll end up scanning a very long LRU
list.
The fix is to scan flush_list and just remove the blocks belonging to
the table from the flush_list, marking them as non-dirty. The blocks
are left in the LRU list for eventual eviction due to aging. The
flush_list is typically much smaller than the LRU list but for cases
where it is very long we have the solution of releasing the
buf_pool->mutex after scanning 1K pages.
buf_page_[set|unset]_sticky(): Use new IO-state BUF_IO_PIN to ensure
that a block stays in the flush_list and LRU list when we release
buf_pool->mutex. Previously we have been abusing BUF_IO_READ to achieve
this.
Refactored the test case: hardened and extended it. Created test inc file
to abstract the task of relocating binlogs.
Also, disabled it on windows for not cluttering the test case any further,
as it depends heavily on doing filesystem operations and path handling.
There was memory leak when running some tests on PB2.
The reason of the failure is an early return from change_master()
that was supposed to deallocate a dyn-array.
Fixed with relocating the dyn-array's destructor at ~LEX() that is
the end of the session, per Gleb's patch idea.
Two optimizations were done: the static buffer for the dyn-array to base on,
and the array initialization is called precisely when it's necessary rather than
per each CHANGE-MASTER as before.
