Windows: Fix a bug that the wrong log file is reopened with log rotate setting when flushing or graceful reloading

[daipom]

When log rotation is enabled on Windows, the log path is separated for each process.
But, the new path value is not applied to the instance variable, so it breaks the following behavior, since #2663.

fluentd/lib/fluent/supervisor.rb

Lines 576 to 580 in bf7498d

	if @chuser || @chgroup
	chuid = @chuser ? ServerEngine::Privilege.get_etc_passwd(@chuser).uid : nil
	chgid = @chgroup ? ServerEngine::Privilege.get_etc_group(@chgroup).gid : nil
	File.chown(chuid, chgid, @path)
	end

fluentd/lib/fluent/supervisor.rb

Lines 601 to 606 in bf7498d

	def reopen!
	if @path && @path != "-"
	@logdev.reopen(@path, "a")
	end
	self
	end

I can't assume that File.chown is used on Windows, but reopen must be fixed.
(--user and --group are not intended to be used on Windows, right? If so, I will add some comments to the doc and code.)

Which issue(s) this PR fixes:
I found this problem during fixing:

• Logger: Log format setting is not applied to the first log #4037

I thought this should be fixed in a separate PR, so I made this PR first.

What this PR does / why we need it:
This bug breaks the function of reopening the log file with log rotation enabled on Windows, since

• Support logrotate on Windows correctly. Fix #2446 #2663

We need this fix to fix this problem.

How to Reproduce

• Prepare config

<system>
  <log>
    rotate_age 5
    rotate_size 2000
  </log>
</system>

 <source>
   @type sample
   tag test
 </source>
 
 <match test.**>
   @type stdout
 </match>

• Run Fluentd

$ bundle exec fluentd -c /test/fluent.conf -o /test/log/fluent.log

• We can find the log files

fluent-supervisor-0.log
fluent-0.log
fluent-0.log.0
...

• Flush(SIGUSR1)
  • We can check the pid of the supervisor process in fluent-supervisor-0.log.
  • starting fluentd-1.15.3 pid=4208 ruby="2.6.6"

$ fluent-ctl flush {pid of supervisor process}

• Then Fluentd reopens fluent.log for the worker process, although this should be fluent-0.log.

We can find a similar problem in graceful reloading(SIGUSR2), but this often causes an error in rotating supervisor log file.
This seems to have another bug, I would like to exclude this bug for now.

$ fluent-ctl reload {pid of supervisor process} (another shell)
log shifting failed. Permission denied @ rb_file_s_rename - (.\fluent\log\fluent-supervisor-0.log, .\fluent\log\fluent-supervisor-0.log.0)
log writing failed. closed stream

Docs Changes:
Not needed.

Release Note:
Same as the title.

[daipom]

Oh, I have to add tests.

[ashie]

Good catch!

[ashie]

I'll wait merging this until you add tests for it.

[daipom]

I'll wait merging this until you add tests for it.

Thanks. Please.

[daipom]

I added tests.

I confirmed one of those tests failed on master on Windows and this PR fixed it.

[daipom]

I wonder why using mock.proxy, not just mock.

• https://github.com/fluent/fluentd/pull/4054/files#diff-eddebe316e0a90606efd872b0fcae79fa441e6edba5b3bba0571bef9410368d8R37

When I change as follows, the test fails. I don't understand what the difference is.
What does the proxy being not passed a block mean?

- mock.proxy(File).chmod(0o777, path.parent.to_s).once
+ mock(File).chmod(0o777, path.parent.to_s).once

[ashie]

Probably it intends to check File.chmod is called just once.
And need to call real File.chmod for following checks.

[daipom]

need to call real File.chmod for following checks.

Oh, thanks! Just simple mock(File).chmod destroys the original behavior of File.chmod, so we have to use proxy here.

[ashie]

Thanks!

[daipom]

Thanks for your review!