item_cmpfunc.h includes pcre2.h, so with the bundled pcre2
it has to be built before anything that includes pcre2.h .
And item_cmpfunc.h is indirectly included everywhere,
also in many plugins.

Somehow Ninja and Makefiles generators can still deduce the
correct build dependencies, but Visual Studio generator cannot.

Two changes:
* move pcre2.h from item_cmpfunc.h to item_cmpfunc.cc
* create an explicit dependency on pcre2 for the server

Related changes in the server code.

In commit befde6e9
a bogus condition was added, aiming to avoid debug assertion failures
during ALTER TABLE...IMPORT TABLESPACE.

page_cur_delete_rec(): Remove the bogus condition, and replace the
use of buf_block_modify_clock_inc() with a lower-level operation
that skips the debug checks that would fail during IMPORT.

Replace fil_space_get_flags(space) == ULINT_UNDEFINED
with the functionally equivalent fil_space_get_size(space) == 0.

The debug instrumentation with the MLOG_LSN pseudo-record has not been
used for debugging for years. Let us remove this code now.
It would have to be removed as part of MDEV-12353 or MDEV-14425 anyway,
when implementing a new redo log file format.

One reference to PageConverter::m_page_zip_ptr was not adjusted
commit 745fd4b3 when the
data member was removed.

Apparently, regular expression operations that remove entire lines
of output do not work with list_files, and hence the adjustments in
commit 1c282d4b were ineffective.

For cat_file (preceded by list_files_write_file) the replace_regex
does work.

For some reason, for suite/parts/inc/partition_crash_exchange.inc
some file names will be lost when using list_files_write_file
instead of list_files.

We use a precise pattern match. dict_mem_create_temporary_tablename()
is generating #sql-ib names followed by decimal digits only.

This commit updates the wsrep-lib. The changes are a cleanup in
client_state TOI processing and stub methods for future extensions.

Pass buf_block_t* to more functions that write redo log.

page_zip_write_node_ptr(), page_zip_write_blob_ptr(),
page_zip_compress_write_log_no_data():
Take buf_block_t* as parameter, and do not tolerate mtr=NULL.

page_zip_compress(): Do not tolerate mtr=NULL.

page_zip_dir_insert(): Take page_cur_t* as parameter.

mlog_write_initial_log_record(): Remove. This function was unused.

RecIterator::remove(): Remove the redundant page_zip parameter.

PageConverter::m_page_zip_ptr: Remove.

page_cur_insert_rec_low(): Take page_cur_t* as a parameter,
and do not tolerate mtr=NULL.

page_cur_insert_rec_zip(): Do not tolerate mtr=NULL.

page_cur_delete_rec(): Do not tolerate mtr=NULL.

page_delete_rec(): Merge with the only caller, RecIterator::remove().

RecIterator::m_mtr: New data member: a dummy mini-transaction.

The adjustments in adb117cf
turned out to be insufficient.

We disable the MDEV-21189 test galera.galera_partition
because it times out.

The type handler and unsigned_flag erroneously went out of sync in
Item_func_minus::fix_length_and_dec.

offset_t: this is a type which represents one record offset.
It's unsigned short int.

a lot of functions: replace ulint with offset_t

btr_pcur_restore_position_func(),
page_validate(),
row_ins_scan_sec_index_for_duplicate(),
row_upd_clust_rec_by_insert_inherit_func(),
row_vers_impl_x_locked_low(),
trx_undo_prev_version_build():
  allocate record offsets on the stack instead of waiting for rec_get_offsets()
  to allocate it from mem_heap_t. So, reducing  memory allocations.

RECORD_OFFSET, INDEX_OFFSET:
  now it's less convenient to store pointers in offset_t*
  array. One pointer occupies now several offset_t. And those constant are start
  indexes into array to places where to store pointer values

REC_OFFS_HEADER_SIZE: adjusted for the new reality

REC_OFFS_NORMAL_SIZE:
  increase size from 100 to 300 which means less heap allocations.
  And sizeof(offset_t[REC_OFFS_NORMAL_SIZE]) now is 600 bytes which
  is smaller than previous 800 bytes.

REC_OFFS_SEC_INDEX_SIZE: adjusted for the new reality

rem0rec.h, rem0rec.ic, rem0rec.cc:
  various arguments, return values and local variables types were changed to
  fix numerous integer conversions issues.

enum field_type_t:
  offset types concept was introduces which replaces old offset flags stuff.
  Like in earlier version, 2 upper bits are used to store offset type.
  And this enum represents those types.

REC_OFFS_SQL_NULL, REC_OFFS_MASK: removed

get_type(), set_type(), get_value(), combine():
  these are convenience functions to work with offsets and it's types

rec_offs_base()[0]:
  still uses an old scheme with flags REC_OFFS_COMPACT and REC_OFFS_EXTERNAL

rec_offs_base()[i]:
  these have type offset_t now. Two upper bits contains type.

recv_dblwr_t::list is used for appending to the beginning and iterating
through its elements. std::deque fits better for that purpose because
it does less allocations than std::forward_list and provides better memory
locality.

When backpatching a forward GOTO label, the old code erroneously
used CURSOR/HANDLER difference between context frames "c" and "a" to tune
a cpop/hpop command. So the cpop/hpop command later tried to pop
all cursors/handlers declared between "a" and "c", but those between
"b" and "c" were not cpushed/hpoped yet during the execution of "GOTO x".

Fixing the code to use the difference between frames "b" and "a" only.

BEGIN     -- a
 ...
GOTO x;   -- b
 ...
<<x>>     -- c
 ...
END       -- d

fil_node_t::find_metadata() tries to find out whether file
is on an SSD, and the disk sector size.
On Windows, it opens the corresponding volume for finding this data.

This does not go well, if datadir is on network path/UNC. The volume name
is invalid, CreateFile() function fails, and a cryptic (from the end user
perspective) error is reported. Like this

[ERROR] InnoDB: File \\.\\\workpc\work: 'CreateFile()' returned OS error 203.

The fix is not to report error if open volume failed, and the path was not
on fixed disk, i.e not on HDD or SSD. This is not a fatal error, there is
a fallback anyway.

copy thread)

mariabackup hangs waiting until innodb redo log thread read log till certain
LSN, and it waits under FTWRL. If there is redo log read error in the thread,
it is finished, and main thread knows nothing about it, what leads to hanging.
As it hangs under FTWRL, slave threads on server side can be blocked due
to MDL lock conflict.

The fix is to finish mariabackup with error message on innodb redo log read
failure.

The issue here is window function makes the passed string object
to point to an area in a temporary table's record buffer.
Then, the temporary table is freed, together with its record buffer.
Then, Item_cache_str attempts to read this value.

The fix is to call value_buff.copy(). This will make the value_buff to store
its string in a buffer that it owns, which will not disappear unexpectedly.

Use std::queue backed by std::deque instead of list because it does less
allocations which still having O(1) push and pop operations.

Also store trx_purge_rec_t directly, because its only 16 bytes and allocating
it is to wasteful. This should be faster.

purge_node_t::purge_node_t: container is already empty after creation

purge_node_t::end(): replace clearing container with assertion that it's clear

Simplify loading messages into mariabackup. Do the same as server does
We're forcing english, so there is no attempt to load errmsg.sys

Test galera_sr_ddl_master would sometimes fail due to leftover
streaming replication fragments. Rollbacker thread would attempt to
open streaming_log table to remove the fragments, but would fail in
check_stack_overrun(). Ultimately the check_stack_overrun() failure
was caused by rollbacker missing to switch the victim's THD thread
stack to rollbacker's thread stack.

Also in this patch:
- Remove duplicate functionality in rollbacker helper functions,
  and extract rollbacker fragment removal into function
  wsrep_remove_streaming_fragments()
- Reuse open_for_write() in wsrep_schema::remove_fragments
- Partially revert changes to galera_sr_ddl_master test from
  commit 44a11a7c. Removed unnecessary
  wait condition and isolation level setting

When using huge pages, the innodb_buffer_pool_size cannot necessarily
be restored. Simplify things by restarting the server.

InnoDB RNG maintains global state, causing otherwise unnecessary bus
traffic. Even worse, this is cross-mutex traffic. That is, different
mutexes suffer from contention.

Fixed delay of 4 was verified to give best throughput by OLTP update
index and read-write benchmarks on Intel Broadwell (2/20/40) and
ARM (1/46/46).

This is a backport of ce047900 from
MariaDB Server 10.3.