Correctly handle pathnames by closing opened files again by haslo · Pull Request #1090 · prawnpdf/prawn

haslo · 2018-11-23T14:20:33Z

This PR fixes #975 by using a standard File.open block for handling path names as images.

The io is yielded within that block instead of returned from the method, and thus is correctly closed after it has been used to read from.

pointlessone · 2018-11-23T14:48:27Z

Wouldn't it be easier to change verify_and_open_image to verify_and_read_image that would just return read data, instead of io? I imagine, the overall code would be simpler and easier to read as well.

Also, please add specs.

haslo · 2018-11-23T15:43:50Z

@pointlessone Thanks a lot for the feedback! I absolutely agree, directly reading the content is a lot simpler.

Do you know of a way to test open file handles from inside Ruby, without resorting to OS tools? I've added a file double way of testing whether the block syntax is used for File.open, and a version that uses lsof and only runs on systems that use that, but I'm not perfectly happy with either...

pointlessone · 2018-11-23T15:46:19Z

      raise ArgumentError, "#{io_or_path} not found" unless io_or_path.file?
-      io = io_or_path.open('rb')
-      io
+      File.open(io_or_path, 'rb') do |file_io|


File.read(io_or_path)

Instead of the block open would work as well.

It would have to be #binread though, right?

Yes, you're right. That is closer to the original code.

because Pathname supports a lot of class-level methods from File you could also just call binread on the Pathname instance directly

Suggested change

File.open(io_or_path, 'rb') do |file_io|

io_or_path.binread

You could then simplify your no-files-left-open spec by passing a known-non-existent file path in plus a correctly configured test double/spy and verify that the spy had binread.

Something like this:

it 'uses binread, which closes files opened from pathnames' do pathname_double = instance_double("Pathname", binread: "1234567") expect(Pathname).to receive(:new).with(filename) { pathname_double } _info = pdf.image(filename) expect(pathname_double).to have_received(:binread) end

Not 100% happy with that, but hopefully you see the point - use of a non-existent path ensures that file system accesses outside calling #binread on the double would fail with a system error, so you don't need to verify through something like ObjectSpace.

pointlessone · 2018-11-23T15:50:19Z

Well, this issue was never considered a serious one because Ruby is quite good a closing files for you. And does so fairly quickly.

It does so when it GCs the file object. I imagine, you can write a reliable spec by preventing that from Happening. Let say, you can disable GC for one spec so that opened file object could live long enough for you to do all the assertion you want.

haslo · 2018-11-23T15:56:45Z

Unfortunately, it happens to be serious for us because the files we open live on an SMB share, who complains about open files fairly quickly. The processes using Prawn are Rails workers in Passenger, who doesn't appear to close the files until the process dies and gets replaced by a new one, which can be minutes...

Both specs are reliable, but I'd prefer to just ask File whether it knows of any open handles. It doesn't look like Ruby has anything of that sort though.

pointlessone · 2018-11-23T16:05:54Z

No, Ruby doesn't keep a list of all open files.

You may be observing this more often in Passenger because many web-servers tune GC to not happen during request. However, it should happen in between requests so I don't know why it has to die to free up file handles.

Additionally, the files might be retained by something other than requests and so, leaking.

haslo · 2018-11-23T16:10:54Z

Leaking is of course possible, but the issue arises in the places where we feed Prawn just pathnames as strings. I'm not sure how long it's being retained, the fact is that it's sometimes minutes. We tried to fix the symptoms by always passing file instances instead of filenames (and then closing those ourselves), but figured that's silly when we can just help fixing the cause instead of digging through all our code and maintaining lists of files in dozens of places...

haslo · 2018-11-23T16:28:17Z

I found a way to write the spec that I'm happy with. What do you think, @pointlessone?

pointlessone · 2018-11-23T16:34:18Z

Isn't ObjectSpace MRI-specific? I'd like to keep whatever JRuby compatibility we have.

haslo · 2018-11-23T16:42:37Z

I wasn't aware. What do you think of running that particular spec only for MRI?

pointlessone · 2018-11-23T16:44:41Z

I'm not opposed to the idea of impl-specific specs. As long as there are corresponding specs for other implementations.

haslo · 2018-11-23T16:58:11Z

I've rebased it all into one commit. Thanks again for your feedback!

pointlessone · 2018-11-23T16:59:54Z

+  it 'closes opened files again after getting Pathname objects' do
+    if RUBY_PLATFORM == 'java'
+      # JRuby doesn't support ObjectSpace
+      expect(RUBY_PLATFORM).to eq('java')


There needs to be a spec here. Otherwise we can't be sure it works on Ruby.

I'll try to come up with something... The spec that used lsof isn't really great either, that doesn't test anything for Windows platforms. Or would that work for you?

Right now we don't run tests on Windows but we don't have any platform-specific code either. Everything is Ruby-only so in theory we shouldn't have any major issues with adding Windows. It at all possible I'd prefer it stayed that way.

pointlessone · 2018-11-23T20:10:05Z

    expect(info.height).to eq(453)
  end

+  context 'with File message spy' do


This spec feels a bit too coupled to the implementation and doesn't describe the actual intent particularly well but since I can't give any useful advice on how to improve it, let's keep it. However, tag it with issue. Here's an example.

I agree, I don't like testing the implementation (and only implicitly the logic) either. Something like the version with ObjectSpace but working on every platform would've been awesome...

haslo · 2018-11-26T08:53:33Z

I think it would make sense to add the lsof and the ObjectSpace versions of the specs after all, even if they have no direct equivalents without those tools. The feature the specs test is whether the file handles are closed again. There will then be three specs for that, testing every angle on systems that have all three approaches available, and only testing the implementation-based angle currently in the PR if they don't.

For reference, lsof version:

  it 'closes opened files again after getting Pathname objects' do
    system_has_lsof = system('lsof -v > /dev/null 2>&1')
    system_has_grep = system('grep --version > /dev/null 2>&1')
    if system_has_lsof && system_has_grep
      gc_was_disabled = GC.disable # make sure GC doesn't remove File instance and close file
      open_files_before = `lsof -c ruby | grep "#{filename}"`
      _info = pdf.image(Pathname.new(filename))
      open_files_after = `lsof -c ruby | grep "#{filename}"`
      GC.enable unless gc_was_disabled
      expect(open_files_after).to eq(open_files_before)
    else
      expect(system_has_lsof && system_has_grep).to be_falsy # just for Rubocop
    end
  end

...and ObjectSpace version:

  it 'closes opened files again after getting Pathname objects' do
    if <guard ObjectSpace availability>
      gc_was_disabled = GC.disable # make sure GC doesn't remove File instance and close file
      open_files_before = ObjectSpace.each_object(File).count { |f| !f.closed? }
      _info = pdf.image(Pathname.new(filename))
      open_files_after = ObjectSpace.each_object(File).count { |f| !f.closed? }
      GC.enable unless gc_was_disabled
      expect(open_files_after).to eq(open_files_before)
    else
      expect(<tautology>).to be_truthy # just for Rubocop
    end
  end

If I add these to the PR, even without corresponding specs when the tools aren't available, the logic is tested from three non-perfect angles instead of just one:

Cross platform compatible, but implementation-based instead of logic-based
Logic-based, but only works on linux and macOS platforms
Logic-based, but only works on MRI

If anybody has an idea for a logic-based cross-platform spec after all, I'm all ears. But without that, personally I'd feel happier with these three specs and a comment explaining their raison d'être than just one...

pointlessone · 2018-11-26T09:29:05Z

Sounds reasonable.

However, instead of meaningless expectations in else branches please use pending with a good doc string.

fidothe · 2018-11-26T15:45:55Z

+      end
+    end
+
+    it 'closes opened files, spec with lsof' do


An easier approach to these kind of specs is to wrap the definition with the platform check, which means you avoid the sometimes-this-spec-is-pending-but-not-always problem:

describe "situation" do system_has_lsof = system('lsof -v > /dev/null 2>&1') system_has_grep = system('grep --version > /dev/null 2>&1') if system_has_lsof && system_has_grep it 'closes opened files, spec with lsof' do ... end end if RUBY_PLATFORM != 'java' it 'closes opened files, spec with ObjectSpace' do ... end end end

haslo · 2018-11-30T10:35:14Z

The PR is still in the status "changes requested" ... the specs are as fitting as I think is possible now, right? Is there something more I can / should do?

pointlessone

Thank you for your contribution.

c.f. prawnpdf/prawn#1090

haslo force-pushed the close_open_files branch from b6bf792 to b8db82a Compare November 23, 2018 14:40

pointlessone reviewed Nov 23, 2018

View reviewed changes

haslo force-pushed the close_open_files branch from 2680130 to 4c57766 Compare November 23, 2018 16:12

haslo force-pushed the close_open_files branch from f1d2505 to 2c87a2e Compare November 23, 2018 16:29

haslo force-pushed the close_open_files branch from 767bc2b to 6b3fc50 Compare November 23, 2018 16:57

pointlessone requested changes Nov 23, 2018

View reviewed changes

pointlessone reviewed Nov 23, 2018

View reviewed changes

haslo force-pushed the close_open_files branch from 3cf0b13 to e91658b Compare November 23, 2018 20:28

haslo force-pushed the close_open_files branch from d4f0d82 to 7170bed Compare November 26, 2018 09:46

pointlessone reviewed Nov 26, 2018

View reviewed changes

Comment thread spec/prawn/images_spec.rb Outdated

fidothe reviewed Nov 26, 2018

View reviewed changes

Change pathname handling by closing opened files after reading

45cd3b9

haslo force-pushed the close_open_files branch from 2946c13 to 45cd3b9 Compare November 26, 2018 15:56

pointlessone approved these changes Nov 30, 2018

View reviewed changes

pointlessone merged commit ddbcb43 into prawnpdf:master Nov 30, 2018

pointlessone mentioned this pull request Dec 13, 2018

Ensure files opened are also closed prawnpdf/ttfunk#62

Closed

jkburges added a commit to Biteable/ttfunk that referenced this pull request Dec 14, 2018

Make code more consistent

367b38c

c.f. prawnpdf/prawn#1090

This was referenced Feb 29, 2020

Should draw_image close the file descriptor? hidakatsuya/prawn-emoji#29

Closed

Close opened image files hidakatsuya/prawn-emoji#31

Merged

hidakatsuya mentioned this pull request Jul 17, 2021

Windows環境でPDFが出力されない thinreports/thinreports#8

Closed

pointlessone pushed a commit to prawnpdf/ttfunk that referenced this pull request Dec 14, 2023

Make code more consistent

be94ad8

c.f. prawnpdf/prawn#1090

Uh oh!

Conversation

haslo commented Nov 23, 2018

Uh oh!

pointlessone commented Nov 23, 2018

Uh oh!

haslo commented Nov 23, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

fidothe Nov 23, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pointlessone commented Nov 23, 2018

Uh oh!

haslo commented Nov 23, 2018

Uh oh!

pointlessone commented Nov 23, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

haslo commented Nov 23, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

haslo commented Nov 23, 2018

Uh oh!

pointlessone commented Nov 23, 2018

Uh oh!

haslo commented Nov 23, 2018

Uh oh!

pointlessone commented Nov 23, 2018

Uh oh!

haslo commented Nov 23, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

haslo Nov 23, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

haslo commented Nov 26, 2018

Uh oh!

pointlessone commented Nov 26, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

haslo commented Nov 30, 2018

Uh oh!

pointlessone left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

3 participants

haslo commented Nov 23, 2018 •

edited

Loading

fidothe Nov 23, 2018 •

edited

Loading

pointlessone commented Nov 23, 2018 •

edited

Loading

haslo commented Nov 23, 2018 •

edited

Loading

haslo Nov 23, 2018 •

edited

Loading

pointlessone commented Nov 26, 2018 •

edited

Loading