Fix logger initialization

The logic of initializing logger is so difficult and has caused
many bugs.

This fix simplifies it and prevents logger from outputting some
initial log messages with some settings not applied, such as
`format`.

This fix consists of the following 2 points.

* All logger parameters are now set in the first initialization.
  * Previously, only parameters related to rotation were applied
    first, but it is preferable to apply all parameters on that
    point.
* Remove LoggerInitializer
  * This class was a source of confusion because its role is
    difficult to understand.

TODO: Fix tests.

Signed-off-by: Daijiro Fukuda <[email protected]>

lib/fluent/log.rb

@@ -67,8 +67,29 @@ def self.event_tags
       LEVEL_TEXT.map{|t| "#{LOG_EVENT_TAG_PREFIX}.#{t}" }
     end
 
+    # Create a unique path for each process.
+    #
+    # >>> per_process_path("C:/tmp/test.log", :worker, 1)
+    # C:/tmp/test-1.log
+    # >>> per_process_path("C:/tmp/test.log", :supervisor, 0)
+    # C:/tmp/test-supervisor-0.log
+    def self.per_process_path(path, process_type, worker_id)
+      path = Pathname(path)
+      ext = path.extname
+
+      if process_type == :supervisor
+        suffix = "-#{process_type}-0#{ext}"  # "-0" for backword compatibility.
+      else
+        suffix = "-#{worker_id}#{ext}"
+      end
+      return path.sub_ext(suffix).to_s
+    end
+
     def initialize(logger, opts={})
-      # overwrites logger.level= so that config reloading resets level of Fluentd::Log
+      # When ServerEngine changes the logger.level, the Fluentd logger level should also change.
+      # So overwrites logger.level= below.
+      # However, currently Fluentd doesn't use the ServerEngine's reloading feature,
+      # so maybe we don't need this overwriting anymore.
       orig_logger_level_setter = logger.class.public_instance_method(:level=).bind(logger)
       me = self
       # The original ruby logger sets the number as each log level like below.
@@ -92,6 +113,7 @@ def initialize(logger, opts={})
       # So if serverengine's logger level is changed, fluentd's log level will be changed to that + 1.
       logger.define_singleton_method(:level=) {|level| orig_logger_level_setter.call(level); me.level = self.level + 1 }
 
+      @path = opts[:path]
       @logger = logger
       @out = logger.instance_variable_get(:@logdev)
       @level = logger.level + 1
@@ -102,7 +124,8 @@ def initialize(logger, opts={})
       @time_format = nil
       @formatter = nil
 
-      self.format = :text
+      self.format = opts.fetch(:format, :text)
+      self.time_format = opts[:time_format] if opts.key?(:time_format)
       enable_color out.tty?
       # TODO: This variable name is unclear so we should change to better name.
       @threads_exclude_events = []
@@ -156,6 +179,10 @@ def dup
     attr_reader :time_format
     attr_accessor :log_event_enabled, :ignore_repeated_log_interval, :ignore_same_log_interval, :suppress_repeated_stacktrace
     attr_accessor :out
+    # Strictly speaking, we should also change @logger.level when the setter of @level is called.
+    # Currently, we don't need to do it, since Fluentd::Log doesn't use ServerEngine::DaemonLogger.level.
+    # Since We overwrites logger.level= so that @logger.level is applied to @level,
+    # we need to find a good way to do this, otherwise we will end up in an endless loop.
     attr_accessor :level
     attr_accessor :optional_header, :optional_attrs
 
@@ -202,8 +229,12 @@ def time_format=(time_fmt)
       @time_formatter = Strftime.new(@time_format) rescue nil
     end
 
+    def stdout?
+      @out == $stdout
+    end
+
     def reopen!
-      @logger.reopen! if @logger
+      @out.reopen(@path, "a") if @path && @path != "-"
       nil
     end