Revert the XtraDB changes, because 10.2 does not currently build with XtraDB.
Also omit some changes that need further investigation.
Ensure that all callers of partition_info::get_clone() are passing this!=NULL.
buf_block_init(): Initialize buf_page_t::flush_type.
For some reason, Valgrind 3.12.0 would seem to flag some
bits in adjacent bitfields as uninitialized, even though only
the two bits of flush_type were left uninitialized. Initialize
the field to get rid of many warnings.
buf_page_init_low(): Initialize buf_page_t::old.
For some reason, Valgrind 3.12.0 would seem to flag all 32
bits uninitialized when buf_page_init_for_read() invokes
buf_LRU_add_block(bpage, TRUE). This would trigger bogus warnings
for buf_page_t::freed_page_clock being uninitialized.
(The V-bits would later claim that only "old" is initialized
in the 32-bit word.) Perhaps recent compilers
(GCC 6.2.1 and clang 4.0.0) generate more optimized x86_64 code
for bitfield operations, confusing Valgrind?
mach_write_to_1(), mach_write_to_2(), mach_write_to_3():
Rewrite the assertions that ensure that the most significant
bits are zero. Apparently, clang 4.0.0 would optimize expressions
of the form ((n | 0xFF) <= 0x100) to (n <= 0x100). The redundant
0xFF was added in the first place in order to suppress a
Valgrind warning. (Valgrind would warn about comparing uninitialized
values even in the case when the uninitialized bits do not affect
the result of the comparison.)
In functions that declare pointer parameters as nonnull,
remove nullness checks, because GCC would optimize them away anyway.
Use #ifdef instead of #if when checking for a configuration flag.
Clang says that left shifts of negative values are undefined.
So, use ~0U instead of ~0 in a number of macros.
Some functions that were defined as UNIV_INLINE were declared as
UNIV_INTERN. Consistently use the same type of linkage.
ibuf_merge_or_delete_for_page() could pass bitmap_page=NULL to
buf_page_print(), conflicting with the __attribute__((nonnull)).
Threads may fall asleep forever while acquiring InnoDB rw-lock on Power8. This
regression was introduced along with InnoDB atomic operations fixes.
The problem was that proper memory order wasn't enforced between "writers"
store and "lock_word" load:
my_atomic_store32((int32*) &lock->waiters, 1);
...
local_lock_word = lock->lock_word;
Locking protocol is such that store to "writers" must be completed before load
of "lock_word". my_atomic_store32() was expected to enforce memory order because
it issues strongest MY_MEMORY_ORDER_SEQ_CST memory barrier.
According to C11:
- any operation with this memory order is both an acquire operation and a
release operation
- for atomic store order must be one of memory_order_relaxed
memory_order_release or memory_order_seq_cst. Otherwise the behavior is
undefined.
That is it doesn't say explicitly that this expectation is wrong, but there are
indications that acquire (which is actually supposed to guarantee memory order
in this case) may not be issued along with MY_MEMORY_ORDER_SEQ_CST.
A good fix for this is to encode waiters into lock_word, but it is a bit too
intrusive. Instead we change atomic store to atomic fetch-and-store, which
does memory load and is guaranteed to issue acquire.
Simplify away recursive flag: it is not necessary for rw-locks to operate
properly. Now writer_thread != 0 means recursive.
As we only need correct value of writer_thread only in writer_thread itself
it is rather safe to load and update it non-atomically.
This patch also fixes potential reorder of "writer_thread" and "recursive"
loads (aka MDEV-7148), which was reopened along with InnoDB thread fences
simplification. Previous versions are unaffected, because they have os_rmb
in rw_lock_lock_word_decr(). It wasn't observed at the moment of writing
though.
Due to high memory reqirements spider tests fail often on automated testing
VM's due to rather limited resource allocation.
For example with 10.2 spider needs at least 200M * 8 mysqld instances = 1.6Gb
RAM per mtr instance. With --parallel=4 it needs 6.4Gb, while appropriate hosts
have just 3Gb.
The crash is caused due to a race condition where wsrep
startup threads invoke ha_maria::implicit_commit() method
while maria_hton is partially initialized. The fix is to
skip this method if plugins are uninitialized.
With "WL#6047 - Do not allocate trx id for read-only transactions"
m_prebuilt->trx->id is always 0 for read-only transactions. This makes
it useless as an index for fuzzy counters.
Use server thread id instead similarly to MySQL.
for InnoDB tables"
Don't use thr_lock.c locks for InnoDB tables. Below is list of changes that
were needed to implement this:
- HANDLER OPEN acquireis MDL_SHARED_READ instead of MDL_SHARED
- HANDLER READ calls external_lock() even if SE is not going to be locked by
THR_LOCK
- InnoDB lock wait timeouts are now honored which are much shorter by default
than server lock wait timeouts (1 year vs 50 seconds)
- with @@autocommit= 1 LOCK TABLES disables autocommit implicitely, though
user still sees @@autocommt= 1
- the above starts implicit transaction
- transactions started by LOCK TABLES are now rolled back on disconnect
(previously everything was committed due to autocommit)
- transactions started by LOCK TABLES are now rolled back by ROLLBACK
(previously everything was committed due to autocommit)
- it is now impossible to change BINLOG_FORMAT under LOCK TABLES (at least
to statement) due to running transaction
- LOCK TABLES WRITE is additionally handled by MDL
- ...in contrast LOCK TABLES READ protection against DML is pure InnoDB
- combining transactional and non-transactional tables under LOCK TABLES
may cause rolled back changes in transactional table and "committed"
changes in non-transactional table
- user may disable innodb_table_locks, which will cause LOCK TABLES to be
noop basically
Removed tests for BUG#45143 and BUG#55930 which cover InnoDB + THR_LOCK. To
operate properly these tests require code flow to go through THR_LOCK debug
sync points, which is not the case after this patch. These tests are removed
by WL#6671 as well. An alternative is to port them to different storage engine.
Issue:
======
Currently the approach we take to find the chunk corresponding to a given
pointer uses srv_buf_pool_chunk_unit based on the assumption that
srv_buf_pool_chunk_unit is the total size of all pages in a buffer pool
chunk. We first step back by srv_buf_pool_chunk_unit bytes and use
std::map::upper_bound() to find the first chunk in the map whose key >= the
resulting pointer.
However, the real size of a chunk (and thus, the total size of its pages)
may differ from the value configured with innodb_buffer_pool_chunk_size
due to rounding up to the OS page size. So, in some cases the above logic
gives us the wrong chunk.
Fix:
====
We find out the chunk corresponding to the give pointer without using
srv_buf_pool_chunk_unit. This is done by using std::map::upper_bound()
to find the next chunk in the map which appears right after the pointer and
decrementing the iterator, which would give us the chunk the pointer
belongs to.
Contribution by Alexey Kopytov.
RB: 13347
Reviewed-by: Debarun Banerjee <debarun.banerjee@oracle.com>
No -DHAVE_LIBNUMA=1 was passed to the source compile (and the
global include/my_config.h wasn't used).
This also is Linux only so corrected the cmake macro.
Fixed indenting in cmake macro.
Removed NUMA defination from include/my_config.h as its only
in the storage engine.
Thanks Elena Stepanova and Vladislav Vaintroub for the detailed
list of bugs/questions.
Signed-off-by: Daniel Black <daniel.black@au.ibm.com>
Clean-up nolock.h: it doesn't serve any purpose anymore. Appropriate code moved
to x86-gcc.h and my_atomic.h.
If gcc sync bultins were detected, we want to make use of them independently of
__GNUC__ definition. E.g. XLC simulates those, but doesn't define __GNUC__.
HS/Spider: According to AIX manual alloca() returns char*, which cannot be
casted to any type with static_cast. Use explicit cast instead.
MDL: Removed namemangling pragma, which didn't let MariaDB build with XLC.
WSREP: _int64 seem to be conflicting name with XLC, replaced with _integer64.
CONNECT: RTLD_NOLOAD is GNU extention. Removed rather meaningless check if
library is loaded. Multiple dlopen()'s of the same library are permitted,
and it never gets closed anyway. Except for error, which was a bug: it may
close library, which can still be referenced by other subsystems.
InnoDB: __ppc_get_timebase() is GNU extention. Only use it when __GLIBC__ is
defined.
Based on contribution by flynn1973.
For Debian the rules file is the main makefile and assuming that a upstream
makefile will mangle the Debian packaging files creates false alerts
from static analysis tools and other problems.
On PPC64 high-loaded server may crash due to assertion failure in InnoDB
rwlocks code.
This happened because load order between "recursive" and "writer_thread"
wasn't properly enforced.
Clean-up periodic mutex/rwlock waiters wake up. This was a hack needed to
workaround broken mutexes/rwlocks implementation. We must have sane
implementations now and don't need these anymore: release thread is
guaranteed to wake up waiters.
Removed redundant ifdef that has equivalent code in both branches.
Removed os0atomic.h and os0atomic.ic: not used anymore.
Clean-up unused cmake checks.
No point to issue RELEASE memory barrier in os_thread_create_func(): thread
creation is full memory barrier.
No point to issue os_wmb in rw_lock_set_waiter_flag() and
rw_lock_reset_waiter_flag(): this is deadcode and it is unlikely operational
anyway. If atomic builtins are unavailable - memory barriers are most certainly
unavailable too.
RELEASE memory barrier is definitely abused in buf_pool_withdraw_blocks(): most
probably it was supposed to commit volatile variable update, which is not what
memory barriers actually do. To operate properly it needs corresponding ACQUIRE
barrier without an associated atomic operation anyway.
ACQUIRE memory barrier is definitely abused in log_write_up_to(): most probably
it was supposed to synchronize dirty read of log_sys->write_lsn. To operate
properly it needs corresponding RELEASE barrier without an associated atomic
operation anyway.
Removed a bunch of ACQUIRE memory barriers from InnoDB rwlocks. They're
meaningless without corresponding RELEASE memory barriers.
Valid usage example of memory barriers without an associated atomic operation:
http://en.cppreference.com/w/cpp/atomic/atomic_thread_fence
Replaced InnoDB atomic operations with server atomic operations.
Moved INNODB_RW_LOCKS_USE_ATOMICS - it is always defined (code won't compile
otherwise).
NOTE: InnoDB uses thread identifiers as a target for atomic operations.
Thread identifiers should be considered opaque: any attempt to use a
thread ID other than in pthreads calls is nonportable and can lead to
unspecified results.
