Fixed by doing adding external_lock(F_UNLCK) in free_locks.
I also moved unlock of cache_handler to be after changing lock type of
cached tables to ensure that no one can use cached table while this is
happening (as cache table is locked with write lock).
I also fixed a wrong mutex order bug in ha_cache::flush_insert_cache()

Other things:
- Addded share::cached_rows to track inserted rows. This is only used
  for asserts to check the number of rows in the cache.
- Fixed wrong mysql_file_chsize() in case of repair

Other things:
- Added test from columnstore team
- Fixed two reported bugs from columnstore team
- Call free_locks as part of start_trans() instead of get_status()
  to ensure that we have locks both for cached table and cache table
  before we try to free any.
- Store pointers to lock->get_status and lock->update_status for the
  cached table. Was needed by ha_tina in flush_insert_cache to make
  new insert rows visible for the SELECT that caused the flush

- The THR_LOCK org_lock must be stored in a shared structure so that all
  instances of a table can use it. Fixed by adding a ha_cache_share object
  that keeps track of this one.
- Fixed wrong test in get_status_and_flush_cache to detect in insert command
- Fixed in get_status_and_flush_cache that we always free the insert lock
  if we don't need it.

This is implemented on top of csv engine. Should be moved to ColumnStore
ASAP

Drop Aliases from the service file directive. Aliases in the service
file only take effect when the service is enabled. This is the case
because Aliases in service files do not have to be unique across various
services.

Shipping symlinks guarantees that one can always enable mariadb service
using `systemctl enable mysql` or `systemctl enable mysqld` and makes
the commands indempotent.

Post change of 8447edb7, --plugin-dir
is set in the default my.cnf file under [client]. The value is automatically
detected during mtr startup.

MDEV-20502 Queries against spider tables return wrong values for columns following constant declarations.

When executing a query like "select id, 0 as const, val from ...", there are 3 columns(items) in Query->select at handlerton->create_group_by(). After that, MariaDB makes a temporary table with 2 columns. The skipped items are const item, so fixing Spider to skip const items for items at Query->select.

- On Windows, do not handle lack of SeLockMemory privilege as fatal error.
Just like on any other platform, there is a fallback to ordinary pages.
It is better than server that does not start, silently.

- On Windows, remove incorrect  irritating "fallback to conventional pages failed"
from the warning, when allocating large pages fails.

The test innodb.innodb_wl6326 that had been disabled in 10.4 due to
MDEV-21535 is failing on 10.5 due to a different reason: the removal
of the MLOG_COMP_END_COPY_CREATED operations in MDEV-12353
commit 276f996a caused PAGE_LAST_INSERT
to be set to something nonzero by the function page_copy_rec_list_end().

This in turn would cause btr_page_get_split_rec_to_right() to behave
differently: we would not attempt to split the page at all, but simply
insert the new record into the new, empty, right leaf page.

Even though the change reduced the sizes of some tables, it is better
to aim for balanced trees.

page_copy_rec_list_end(), PageBulk::finishPage():
Preserve PAGE_LAST_INSERT, PAGE_N_DIRECTION, PAGE_DIRECTION.

PageBulk::finish(): Move some common code from PageBulk::finishPage().

Without this, the mtr will fail to load client plugins and the following
tests would fail:

- main.plugin_auth_qa_2
- main.plugin_auth_qa_3
- main.plugin_auth
- plugins.multiauth
- plugins.auth_ed25519

When testing installed packages, the client (mariadb or mysql) can have
plugins stored in a separate location compared to server. This change
ensures that the correct paths are used for the plugin-dir directive in
the my.cnf generated by mysql-test-run.

The directives in my.cnf will be replaced like so:

[client]
plugin-dir=$client_plugindir

[mysqld]
plugin-dir=$plugindir

where $<variable-name> corresponds to the variable name in mtr.

plugin-dir is provided via --defaults-file=<config-file> already part of
the exec call.

As this was missing, the libmariadb/install.cmake section for Debian
builds was not triggered at all, and existing fixes to make the
Debian paths correct did not apply.

Update libmariadb3 paths in Debian packaging after MariaDB Connector C
3.1.8 upgrade

Also now include caching_sha2_password.so which was previously omitted.

Keep old libmariadbclient.a as symbolic link for backwards compatibility.

Closes: CONC-304
Related: CONC-456, MDEV-22150

MDEV-22218 InnoDB: Failing assertion: node->pcur->rel_pos == BTR_PCUR_ON upon LOAD DATA with NO_BACKSLASH_ESCAPES in SQL_MODE and unique blob in table

`inited == NONE` at the initialization time does not always mean
that it'll be `NONE` later, at the execution time. Use a more complex
caller-specific logic to decide whether to create a cloned lookup handler.

Besides LOAD (as in the original bug report) make sure that all
prepare_for_insert() invocations are covered by tests. Add tests for
CREATE ... SELECT, multi-UPDATE, and multi-DELETE.

Don't enable write cache with long uniques.

CMake parameter

This includes compiler-specific dependencies
- tiny part of CRT, mostly memcpy/memset in msvcr140.dll
- exception handling in msvc140_1.dll
- C++ standard library in msvcp140
not much else

CMake is smart enough to copy the dependencies into the bin directory
itself.

Stop linking plugins to the server executable on Windows.
Instead, extract whole server functionality into a large DLL, called
server.dll. Link both plugins, and small server "stub" exe to it.

This eliminates plugin dependency on the name of the server executable.
It also reduces the size of the packages (since tiny mysqld.exe
and mariadbd.exe are now both linked to one big DLL)

Also, simplify the functionality of exporing all symbols from selected
static libraries. Rely on WINDOWS_EXPORT_ALL_SYMBOLS, rather than old
self-backed solution.

fix compile error

replace GetProcAddress(GetModuleHandle(NULL), "variable_name")
for server exported data with actual variable names.

Runtime loading was never required,was error prone
, since symbols could be missing at runtime, and now it actually failed,
because we do not export symbols from executable anymore, but from a shared
library

This did require a MYSQL_PLUGIN_IMPORT decoration for the plugin,
but made the code more straightforward, and avoids missing symbols at
runtime (as mentioned before).

The audit plugin is still doing some dynamic loading, as it aims to work
cross-version. Now it won't work cross-version on Windows, as it already
uses some symbols that are *not* dynamically loaded, e.g fn_format
and those symbols now exported from server.dll , when earlier they were
exported by mysqld.exe

Windows, fixes for storage engine plugin loading
after various rebranding stuff

Create server.dll containing functionality of the whole server
make mariadbd.exe/mysqld.exe a stub that is only  calling mysqld_main()

fix build

This broke clang-cl, apparently when compiling version resource file with
RC tool, because RC would not aceept -Wxxx at all.

- InnoDB is doing un-necessary redo log page initialisation during
recovery and unnecessary traversal of redo log during last phase.
This patch does the optimization of removing unnecessary redo log page
initialisation and detects the memory exhaust earlier.

overflow happening when we add suffix length to sortlength.
Make sure that sortlength is at max UINT_MAX32

Fixed incorrect pointer reference when table is not available.

On by default, suppress "fallback to conventional memory"
warnings as we're never sure what the test has available.

Leave off Windows test for now. Running without SeLockMemoryPrivilege
is fatal if --large-pages is specified.

Both Windows and MMAP capable implementations fell back to a
non-MEM_LARGE_PAGES/HugeTLB allocation with the large page implementaion
failed. These can can be freed by the corresponding function.

Prior to this, if we fell back to a conventional memory, than will
results in deallocation using munmap/VirtualFree on a memory allocated
using my_malloc_lock. At worst this could succeed and the
my_malloc_lock looses its memory without knowing about it.

The bug was already fixed in
commit 1242eb3d
but the test innodb.innodb_wl6326 had not been enabled.

Changed wording in error messages from MySQL to MariaDB. In
cases where the word server could be used instead it was done.

Tests that have these errors recorded were updated.