innodb_log_file_disabled: A new Boolean settable global parameter,
for disabling the InnoDB redo log. When set, the server is not
crash safe.
innodb_log_group_home_dir: Allow the value to be changed with
SET GLOBAL, as long as the log remains in the same file system
or innodb_log_file_disabled=ON was set.
innodb_log_update(): A common function for implementing
SET GLOBAL innodb_log_file_size, innodb_log_file_disabled,
innodb_log_checkpoint_now, innodb_log_group_home_dir.
log_sys.buf_size_requested: The configured innodb_log_buffer_size.
While the log is disabled, we may set log_sys.buf_size
(the actual size of log_sys.buf) differently.
log_sys.disabled: The current setting of innodb_log_file_disabled.
log_t::append_prepare(): Instead of referring to file_size or
capacity() for mmap=true, always refer to buf_size.
When log_sys.disabled holds, log_sys.buf may be much smaller than
log_sys.file_size. Its size is always reflected by log_sys.buf_size.
log_t::attach(): Handle log_sys.disabled.
log_t::disable(): Implements SET innodb_log_file_disabled=ON.
Even if the log used to be in persistent memory, here we will
set up dummy log_sys.buf and log_sys.flush_buf so that the dummy
writes will appear to use file based writes.
log_t::skip_write_buf(): A dummy log writer implementation that is
used when log_sys.disabled holds and the log is not being resized.
log_t::resize_abort(): When the log remains disabled,
"persist" all the log and advance the group_lock and flush_lock
to the current LSN, just like log_t::disable() does.
log_t::resize_start(): Handle log_sys.disabled, that is,
SET GLOBAL innodb_log_file_disabled=OFF when the redo log had
previously been disabled. If we are on persistent memory,
we will "fake" the dummy log_sys.buf to appear as memory-mapped
as well, so that log_t::resize_write() and log_t::resize_write_buf()
can assume that both log files are of the same type (memory-mapped
or file-based). The dummy log_sys.flush_buf will be stored in
log_sys.resize_flush_buf in that case. When moving from memory-mapped
to a regular log file, we will allocate log_sys.checkpoint_buf.
It will be freed in log_t::close().
log_t::resize_rename(): When innodb_log_group_home_dir
is changed between file systems, handle the non-atomic
replacement of the log file in a special way. For a moment,
a recoverable ib_logfile0 will exist in both locations.
log_t::resize_write(): Detect memory-mapped log by
!resize_log.is_opened(). If the memory-mapped log is being re-enabled,
log_sys.resize_flush_buf may temporarily store the value of a dummy
log_sys.flush_buf. During any log resizing or disabling or enabling,
both buf and resize_buf must appear to be either file-based or
memory-mapped.
log_t::write_buf(), log_write_up_to(): Handle the special case that
disable() executed or resize_start() started re-enabling the log while
we were waiting for flush_lock or write_lock. In the latter case,
log_sys.writer will be changed from log_t::skip_write_buf to
log_writer_resizing during the execution of log_write_up_to().
log_t::persist(): Skip the writes if the log is disabled, that is,
a memory-mapped log is in the process being re-enabled. In this case,
we cannot trust log_sys.file_size, and the pmem_persist() could be
invoked on an invalid address range.
log_resize_acquire_group(): Acquier the group locks for write and flush.
log_resize_release_group(): Release the group locks for write and flush.
log_resize_acquire(): Return whether the group locks were elided.
log_write_and_flush_prepare(), log_write_and_flush(): Protect also the
log_sys.is_mmap() case with write_lock and flush_lock, in order to
prevent a race condition between mtr_t::commit_file() and
log_t::disable().
log_t::persist(): Remove the assertions about not holding write_lock or
flush_lock. We will hold them during DDL operations.