Test also on PHP 7.2+

[Al2Klimov]

No description provided.

[Al2Klimov]

Rationale: Who knows all the upcoming changes to PHP and our obstacles?

[dnsmichi]

Mh if nightlies contain bugs, this will break travis tests on PRs. The benefit of testing the git master against nightlies is then gone. Imho this should be done in a Jenkins job, separate from code commits.

[Al2Klimov]

@dnsmichi Well, if even this PR fails to be tested... then you're damn right.

[Thomas-Gelf]

Hint: add allow_failures -> - php: nightly to your matrix

[dnsmichi]

Mh you should see them imho. This allows to react sooner on possibly removed things, especially if you are really testing the new "hot shit".

[Al2Klimov]

Just testing. Don't panic.

[dnsmichi]

Then edit the PR's title to include "WIP: " as prefix :)

[Al2Klimov]

It seems PHP 7.2 has broken our session reopening introduced in efa23ad.

[Al2Klimov]

1. The tests fail on PHP 7.2+ because of our session reopening.
2. Our session reopening is required as we use session_write_close().

But why don't we close the session just once – at shutdown?

CC @lippserd @Thomas-Gelf @nilmerg

[lippserd]

@Al2Klimov session_start() locks the session file until session_write_close() is called. We close the session immediately after read/write operations to prevent Ajax requests from blocking. Smart eh? 😆

[Al2Klimov]

Gentlemen, it's your turn. B-)

[Thomas-Gelf]

Short note: the former session code looks like it tried to avoid issues with "headers already sent". No idea for what use case, but it looks like the new variant would break that countermeasure. Eventually @lippserd can have a look at this. Filter -> good catch! Static code analysis is not that bad at all 😆.

[Al2Klimov]

@Thomas-Gelf What break? I've tested that stuff w/

class IndexController extends Controller
{
    protected $requiresAuthentication = false;

    public function indexAction()
    {
        $s = Session::getSession();
        $s->write();
        $s->write();
        $s->write();
        var_dump(headers_list());
    }
}

The former session code (efa23ad) was requested by you to "Allow sessions to be reopened" – and it still works.

[Thomas-Gelf]

Nothing is wrong with your code alone, and I'm sure tests are fine. So, it's no break per se, but it changes current behavior. To me the old code looks like it has intentionally been placed there. It told PHP to not send headers on it's own (triggered by session_start) at that early stage. There might be a reason for this, other components might rely on deferred header sending. But you should better discuss this with @lippserd.

[Al2Klimov]

... or w/ @nilmerg the author of efa23ad.

[Thomas-Gelf]

@Al2Klimov: just in case you're wondering, the referenced issue 5510 refers to dev.icinga.org, it's #457 here on GitHub.

[Al2Klimov]

No, I'm not. But thank you anyway.

[Al2Klimov]

Just tested w/

class IndexController extends Controller
{
    protected $requiresAuthentication = false;

    public function indexAction()
    {
        $s = Session::getSession();
        $s->write();
        while (ob_get_level()) {
            ob_end_clean();
        }
        echo 'foobar';
        $s->write();
        var_dump(headers_list());
    }
}

(It doesn't work.)

[Al2Klimov]

... but IMO it doesn't break anything... except write operations at shutdown time. :/

[Thomas-Gelf]

@Al2Klimov: there are too many unrelated (but important) changes in this pull request, we should extract the unrelated ones. Could you please start with cc9bf70 and 4f5b1ed and open dedicated pull requests for them? They are clearly bugs, I'd love to merge them immediately. With less priority the same goes true for b3ab01f.

You could also create a dedicated one enabling 7.2 on travis but allowing it to fail, so we could easily merge to master also that part. Once rebased, all branches would then already show test results for 7.2 without officially "breaking".

[Al2Klimov]

Why "a dedicated one" – and not this one? One nice day we just turn it on (disallow failing).

[Al2Klimov]

#3305, #3306, #3307

[Al2Klimov]

@Thomas-Gelf We don't have to allow Travis to fail – if you give green light on #3213

[Thomas-Gelf]

Perfect, merged all three of them, thank you. Same game again please for f334a9e and cf85e66. Also for f8a71b7, but there please extend the commit message and explain your motivation for this commit. Is our default error_reporting too low? Or too strict on PHP 7.2? What would fail (or not fail) without this change?

[Al2Klimov]

cf85e66? IMO we need this only here.

[Al2Klimov]

#3308, #3309

[Thomas-Gelf]

The only commit that will remain in this pull request is 2d5140a - as it is the only one that fit's the topic. The others fix related issues. To me cf85e66 is clearly a bugfix. Parts of this have been triggered by 91b0e98, one line is older. I'm wondering why this didn't fail earlier, I would have expected also older PHP versions to warn on this.

Next, what was the motivation for 6fc90be (PhpCodeValidityTest)? It does obviously more than php -l, as it also loads the code. Also, lint-alike errors would usually be caught by phpcs.

And the last big beast is the session. That one requires special care. Some weird things in the original implementation derive from locking/performance reasons, as @lippserd pointed out. Not writing when nothing changed always makes sense, it saves a lot of I/O. Still, even with no change from time to time we need to force a write, as otherwise GC would kill our session.

Immediately closing the session on write is even more critical, as the default PHP behavior would have devastating effect on our performance. And what remains last: modules like Nagvis are forced to do weird things with sessions. They shut one session down, start the other one, switch back - none of this works reliably when PHP headers are sent out early. This and related issues are what those INI settings should have taken care of.

To keep things short: there is too much duplicated code in PhpSession72. Please have a look at the write() method, just to name one example. There is one differing line, no need to clone the whole method. Please isolate the tasks causing trouble on 7.2 in the original file into dedicate methods and then override only them.

Thanks,
Thomas

[Al2Klimov]

Now as I've outsourced #3310, this works w/o #3213.

[Thomas-Gelf]

We're getting closer :-)

[Thomas-Gelf]

I cherry-picked c8930f3 to speed things up. Please move the fixes for the Session to a dedicated pull request and address above change requests. Thanks for the good work!

[Al2Klimov]

@Thomas-Gelf As long as Travis fails, please wait until I've rebased the other merged PRs.

[SMillerDev]

This is only waiting for the pr to drop 5.3 then?

[Al2Klimov]

@SMillerDev No, it isn't.