There were two problems: RESET QUERY CACHE took a long time to complete
and other threads were blocked during this time.
The patch does three things:
1 fixes a bug with improper use of test-lock-test_again technique.
AKA Double-Checked Locking is applicable here only in few places.
2 Somewhat improves performance of RESET QUERY CACHE.
Do my_hash_reset() instead of deleting elements one by one. Note
however that the slowdown also happens when inserting into sorted
list of free blocks, should be rewritten using balanced tree.
3 Makes RESET QUERY CACHE non-blocking.
The patch adjusts the locking protocol of the query cache in the
following way: it introduces a flag flush_in_progress, which is
set when Query_cache::flush_cache() is in progress. This call
sets the flag on enter, and then releases the lock. Every other
call is able to acquire the lock, but does nothing if
flush_in_progress is set (as if the query cache is disabled).
The only exception is the concurrent calls to
Query_cache::flush_cache(), that are blocked until the flush is
over. When leaving Query_cache::flush_cache(), the lock is
acquired and the flag is reset, and one thread waiting on
Query_cache::flush_cache() (if any) is notified that it may
proceed.
server to crash".
Crash caused by assertion failure happened when one ran SHOW OPEN TABLES
while concurrently doing DROP TABLE (or RENAME TABLE, CREATE TABLE LIKE
or any other command that takes name-lock) in other connection.
For non-debug version of server problem exposed itself as wrong output
of SHOW OPEN TABLES statement (it was missing name-locked tables).
Finally in 5.1 both debug and non-debug versions simply crashed in
this situation due to NULL-pointer dereference.
This problem was caused by the fact that table placeholders which were
added to table cache in order to obtain name-lock had TABLE_SHARE::table_name
set to 0. Therefore they broke assumption that this member is non-0 for
all tables in table cache which was checked by assert in list_open_tables()
(in 5.1 this function simply relies on it).
The fix simply sets this member for such placeholders to appropriate value
making this assumption true again.
This patch also includes test for similar bug 12212 "Crash that happens
during removing of database name from cache" reappeared in 5.1 as bug 19403.
A date can be represented as an int (like 20060101) and as a string (like
"2006.01.01"). When a DATE/TIME field is compared in one SELECT against both
representations the constant propagation mechanism leads to comparison
of DATE as a string and DATE as an int. In this example it compares 2006 and
20060101 integers. Obviously it fails comparison although they represents the
same date.
Now the Item_bool_func2::fix_length_and_dec() function sets the comparison
context for items being compared. I.e. if items compared as strings the
comparison context is STRING.
The constant propagation mechanism now doesn't mix items used in different
comparison contexts. The context check is done in the
Item_field::equal_fields_propagator() and in the change_cond_ref_to_const()
functions.
Also the better fix for bug 21159 is introduced.
Changed the automake build process :
- ./configure.in
- ./sql/Makefile.am
to compile an instrumented parser for debug=yes or debug=full builds
Changed the (primary) runtime invocation of the parser :
- sql/sql_parse.cc
to generate bison traces in stderr when the DBUG "parser_debug" flag is set.
The problem was that the error handling was using a too-small buffer to
print the error message generated. We fix this by not using a buffer at
all, but by using fprintf() directly. There were also some problems with
the error handling in table dumping that was exposed by this fix that were
also corrected.
