MDEV-21382 use fdatasync() for redo log where appropriate

log_t::files::fdatasync(): syncs only data for every log file

os_file_flush_data()
pfs_os_file_flush_data_func(): syncs only data for a given file

storage/innobase/handler/ha_innodb.cc

@@ -18632,7 +18632,7 @@ checkpoint_now_set(THD*, st_mysql_sys_var*, void*, const void* save)
 			  ? log_sys.append_on_checkpoint->size() : 0)
 		       < log_sys.lsn) {
 			log_make_checkpoint();
-			log_sys.log.fsync();
+			log_sys.log.fdatasync();
 		}
 
 		dberr_t err = fil_write_flushed_lsn(log_sys.lsn);

storage/innobase/include/log0log.h

@@ -565,6 +565,8 @@ struct log_t{
     void write(size_t total_offset, span<byte> buf);
     /** flushes OS page cache for all log files */
     void fsync();
+    /** flushes OS page cache (excluding metadata!) for all log files */
+    void fdatasync();
     /** closes all log files */
     void close_files();
 

storage/innobase/include/os0file.h

@@ -787,6 +787,9 @@ The wrapper functions have the prefix of "innodb_". */
 # define os_file_flush(file)					\
 	pfs_os_file_flush_func(file, __FILE__, __LINE__)
 
+#define os_file_flush_data(file)                                              \
+  pfs_os_file_flush_data_func(file, __FILE__, __LINE__)
+
 # define os_file_rename(key, oldpath, newpath)				\
 	pfs_os_file_rename_func(key, oldpath, newpath, __FILE__, __LINE__)
 
@@ -1032,6 +1035,18 @@ pfs_os_file_flush_func(
 	const char*	src_file,
 	uint		src_line);
 
+/** NOTE! Please use the corresponding macro os_file_flush_data(), not directly
+this function!
+This is the performance schema instrumented wrapper function for
+os_file_flush_data() which flushes only(!) data (excluding metadata) from OS
+page cache of a given file to the disk.
+@param[in]	file		Open file handle
+@param[in]	src_file	file name where func invoked
+@param[in]	src_line	line where the func invoked
+@return true if success */
+bool pfs_os_file_flush_data_func(pfs_os_file_t file, const char *src_file,
+                                 uint src_line);
+
 /** NOTE! Please use the corresponding macro os_file_rename(), not directly
 this function!
 This is the performance schema instrumented wrapper function for

storage/innobase/log/log0log.cc

@@ -659,6 +659,23 @@ void log_t::files::fsync()
   log_sys.flushes.fetch_add(1, std::memory_order_release);
 }
 
+void log_t::files::fdatasync()
+{
+  ut_ad(files.size() == file_names.size());
+
+  log_sys.pending_flushes.fetch_add(1, std::memory_order_acquire);
+  for (auto it= files.begin(), end= files.end(); it != end; ++it)
+  {
+    if (!os_file_flush_data(*it))
+    {
+      const auto idx= std::distance(files.begin(), it);
+      ib::fatal() << "os_file_flush_data(" << file_names[idx] << ") failed";
+    }
+  }
+  log_sys.pending_flushes.fetch_add(1, std::memory_order_release);
+  log_sys.flushes.fetch_add(1, std::memory_order_release);
+}
+
 void log_t::files::close_files()
 {
   for (auto it= files.begin(), end= files.end(); it != end; ++it)
@@ -864,7 +881,7 @@ log_write_flush_to_disk_low()
 	calling os_event_set()! */
 	ut_a(log_sys.n_pending_flushes == 1); /* No other threads here */
 
-	log_sys.log.fsync();
+	log_sys.log.fdatasync();
 
 	log_mutex_enter();
 	log_sys.flushed_to_disk_lsn = log_sys.current_flush_lsn;
@@ -1298,7 +1315,7 @@ void log_write_checkpoint_info(lsn_t end_lsn)
 							   : LOG_CHECKPOINT_1,
 			  {buf, OS_FILE_LOG_BLOCK_SIZE});
 
-	log_sys.log.fsync();
+	log_sys.log.fdatasync();
 
 	log_mutex_enter();
 
@@ -1776,7 +1793,7 @@ logs_empty_and_mark_files_at_shutdown(void)
 
 		/* Ensure that all buffered changes are written to the
 		redo log before fil_close_all_files(). */
-		log_sys.log.fsync();
+		log_sys.log.fdatasync();
 	} else {
 		lsn = srv_start_lsn;
 	}

storage/innobase/os/os0file.cc

@@ -4614,3 +4614,24 @@ os_normalize_path(
 		}
 	}
 }
+
+bool pfs_os_file_flush_data_func(pfs_os_file_t file, const char *src_file,
+                                 uint src_line)
+{
+  PSI_file_locker_state state;
+  struct PSI_file_locker *locker= NULL;
+
+  register_pfs_file_io_begin(&state, locker, file, 0, PSI_FILE_SYNC, src_file,
+                             src_line);
+
+#ifdef _WIN32
+  bool result= os_file_flush_func(file);
+#else
+  bool result= true;
+  if (fdatasync(file) == -1)
+    ib::error() << "fdatasync() errno: " << errno;
+#endif
+
+  register_pfs_file_io_end(locker, 0);
+  return result;
+}