Confirmed, upped the severity of this bug.

According to OP, this problem seems to have started in the past few days - maybe it was introduced in 1.20wmf2?

Are POST Request also affected, where titles= is part of the POST data?

It appears that "very long" here corresponds to "512 or more bytes in the parameter". It does not seem to have anything to do with the number of titles, the total length of all parameters, or the number of characters in the URL-encoded representation.

The problem seems to be more general than the simple example I provided first.

For example, on frwiki, if I try to retrieve the list of templates in Hacker, it orks :
http://fr.wikipedia.org/w/api.php?action=query&prop=templates&format=xml&titles=Hacker

<?xml version="1.0"?>
<api>

<query>
  <pages>
    <page pageid="1830125" ns="0" title="Hacker">
      <templates>
        <tl ns="10" title="Modèle:Autres projets" />
        <tl ns="10" title="Modèle:Homonymie" />
      </templates>
    </page>
  </pages>
</query>

</api>

If I had a tltemplates argument with only "Modèle:Homonymie", it works also :
http://fr.wikipedia.org/w/api.php?action=query&prop=templates&format=xml&tltemplates=Mod%C3%A8le%3AHomonymie&titles=Hacker

<?xml version="1.0"?>
<api>

<query>
  <pages>
    <page pageid="1830125" ns="0" title="Hacker">
      <templates>
        <tl ns="10" title="Modèle:Homonymie" />
      </templates>
    </page>
  </pages>
</query>

</api>

But, when I put all disambiguation templates in the tltemplates argument, the result is wrong (Modèle:Homonymie is not returned as one of the templates in Hacker) :
/w/api.php?action=query&prop=templates&format=xml&tltemplates=Mod%C3%A8le%3AArrondissements%20homonymes%7CMod%C3%A8le%3ABandeau%20standard%20pour%20page%20d'homonymie%7CMod%C3%A8le%3ABatailles%20homonymes%7CMod%C3%A8le%3ACantons%20homonymes%7CMod%C3%A8le%3ACommunes%20fran%C3%A7aises%20homonymes%7CMod%C3%A8le%3AFilms%20homonymes%7CMod%C3%A8le%3AGouvernements%20homonymes%7CMod%C3%A8le%3AGuerres%20homonymes%7CMod%C3%A8le%3AHomonymie%7CMod%C3%A8le%3AHomonymie%20bateau%7CMod%C3%A8le%3AHomonymie%20d'%C3%A9tablissements%20scolaires%20ou%20universitaires%7CMod%C3%A8le%3AHomonymie%20d'%C3%AEles%7CMod%C3%A8le%3AHomonymie%20de%20clubs%20sportifs%7CMod%C3%A8le%3AHomonymie%20de%20comt%C3%A9s%7CMod%C3%A8le%3AHomonymie%20de%20monument%7CMod%C3%A8le%3AHomonymie%20de%20nom%20romain%7CMod%C3%A8le%3AHomonymie%20de%20parti%20politique%7CMod%C3%A8le%3AHomonymie%20de%20route%7CMod%C3%A8le%3AHomonymie%20dynastique%7CMod%C3%A8le%3AHomonymie%20vid%C3%A9oludique%7CMod%C3%A8le%3AHomonymie%20%C3%A9difice%20religieux%7CMod%C3%A8le%3AInternationalisation%7CMod%C3%A8le%3AIsom%C3%A9rie%7CMod%C3%A8le%3AParonymie%7CMod%C3%A8le%3APatronyme%7CMod%C3%A8le%3APatronyme%20basque%7CMod%C3%A8le%3APatronyme%20italien%7CMod%C3%A8le%3APatronymie%7CMod%C3%A8le%3APersonnes%20homonymes%7CMod%C3%A8le%3ASaints%20homonymes%7CMod%C3%A8le%3ATitres%20homonymes%7CMod%C3%A8le%3AToponymie%7CMod%C3%A8le%3AUnit%C3%A9s%20homonymes%7CMod%C3%A8le%3AVilles%20homonymes%7CMod%C3%A8le%3A%C3%89difices%20religieux%20homonymes&titles=Hacker

<?xml version="1.0"?>
<api>

<query>
  <pages>
    <page pageid="1830125" ns="0" title="Hacker" />
  </pages>
</query>

</api>

Link to last example was:
http://fr.wikipedia.org/w/api.php?action=query&prop=templates&format=xml&tltemplates=Mod%C3%A8le%3AArrondissements%20homonymes%7CMod%C3%A8le%3ABandeau%20standard%20pour%20page%20d'homonymie%7CMod%C3%A8le%3ABatailles%20homonymes%7CMod%C3%A8le%3ACantons%20homonymes%7CMod%C3%A8le%3ACommunes%20fran%C3%A7aises%20homonymes%7CMod%C3%A8le%3AFilms%20homonymes%7CMod%C3%A8le%3AGouvernements%20homonymes%7CMod%C3%A8le%3AGuerres%20homonymes%7CMod%C3%A8le%3AHomonymie%7CMod%C3%A8le%3AHomonymie%20bateau%7CMod%C3%A8le%3AHomonymie%20d'%C3%A9tablissements%20scolaires%20ou%20universitaires%7CMod%C3%A8le%3AHomonymie%20d'%C3%AEles%7CMod%C3%A8le%3AHomonymie%20de%20clubs%20sportifs%7CMod%C3%A8le%3AHomonymie%20de%20comt%C3%A9s%7CMod%C3%A8le%3AHomonymie%20de%20monument%7CMod%C3%A8le%3AHomonymie%20de%20nom%20romain%7CMod%C3%A8le%3AHomonymie%20de%20parti%20politique%7CMod%C3%A8le%3AHomonymie%20de%20route%7CMod%C3%A8le%3AHomonymie%20dynastique%7CMod%C3%A8le%3AHomonymie%20vid%C3%A9oludique%7CMod%C3%A8le%3AHomonymie%20%C3%A9difice%20religieux%7CMod%C3%A8le%3AInternationalisation%7CMod%C3%A8le%3AIsom%C3%A9rie%7CMod%C3%A8le%3AParonymie%7CMod%C3%A8le%3APatronyme%7CMod%C3%A8le%3APatronyme%20basque%7CMod%C3%A8le%3APatronyme%20italien%7CMod%C3%A8le%3APatronymie%7CMod%C3%A8le%3APersonnes%20homonymes%7CMod%C3%A8le%3ASaints%20homonymes%7CMod%C3%A8le%3ATitres%20homonymes%7CMod%C3%A8le%3AToponymie%7CMod%C3%A8le%3AUnit%C3%A9s%20homonymes%7CMod%C3%A8le%3AVilles%20homonymes%7CMod%C3%A8le%3A%C3%89difices%20religieux%20homonymes&titles=Hacker

Probably the same underlying problem: it's searching for "Modèle:Homonymie" instead of "Modèle:Homonymie". Trim the value of tltemplates down to 511 bytes or less (before URL-encoding) and it'll probably work. Or try it with a POST, as Umherirrender suggested.

I just tried with POST instead of GET : it works.
So the problem seems to appear only with GET requests.

Bumping to highest as this appears to be a regression with the recent rollout. Notifying Robla and Sumana.

• Bug 36799 has been marked as a duplicate of this bug. ***

I think I might have figured this out.

In a post on enwiki from May 11,[1] we are told that Roan changed the "PCRE recursion limit" from the default 100k to 1k. I assume this is referring to PHP's "pcre.recursion_limit" setting,[2] which indeed has a default of 100000.

One thing the recursion limit affects how often regexes with subexpressions like "(x)+" can match. It seems that each match by "+" there uses up 2 of the recursion limit; with a value of 1024, it can match at most 511 times. If it would match 512 times, preg_match will return false instead. You can test this easily enough if you have a recent-enough command-line PHP:

php -r 'ini_set("pcre.recursion_limit", 1024); var_dump(preg_match("/(x)+/", str_repeat("x", 511)));'
php -r 'ini_set("pcre.recursion_limit", 1024); var_dump(preg_match("/(x)+/", str_repeat("x", 512)));'

The first will succeed, while the second will fail. But if you bump the 1024 to 1026, the second will start working.

So what seems to be going on is this: The API uses the methods in WebRequest to get the parameters from the client, all of which seem to come down to getGPCVal. For any parameter that exists in $_GET (even if overridden by $_POST), getGPCVal passes the value through Language::checkTitleEncoding to make sure it's valid UTF-8. And due to the low recursion limit, the regex in Language::checkTitleEncoding that tries to check whether the value is valid UTF-8 will now think it is ''not'' valid if the value is more than 511 characters long, so it will treat it as the fallback 8-bit encoding (windows-1252 for most languages), which gives the familiar "è" mojibake.

If I'm right, the fix for this bug would be to revert Roan's change to the "pcre.recursion_limit" setting (and fix whatever PageTriage's problem is in some other way), or at least turn it up to something more reasonable than 1024. I'd expect this is causing problems in other areas of the code, too.

[1]: https://en.wikipedia.org/w/index.php?title=Wikipedia_talk:Database_reports&diff=491927371&oldid=491919743
[2]: http://us.php.net/manual/en/pcre.configuration.php#ini.pcre.recursion-limit

sumanah wrote:

Keywording i18n.

(In reply to comment #10)

I'd expect this is causing problems in other areas of the code, too.

And it causing problems for many clients, which using titles= with 50 urlencoded titles. Please use a temp fix soon (reset the limit) and look into the real problem after that.

Thanks.

The regexp recursion limit aside, is using a regexp to check for UTF-8 appropriate? Why not use mb_check_encoding() if available? Other operations in Language.php do make use of the mb_* functions...

http://php.net/manual/en/function.mb-check-encoding.php

• Bug 37021 has been marked as a duplicate of this bug. ***

(In reply to comment #10)

If I'm right, the fix for this bug would be to revert Roan's change to the
"pcre.recursion_limit" setting (and fix whatever PageTriage's problem is in
some other way), or at least turn it up to something more reasonable than 1024.
I'd expect this is causing problems in other areas of the code, too.

Sam, have you had a chance to look at this yet?

(In reply to comment #15)

(In reply to comment #10)

If I'm right, the fix for this bug would be to revert Roan's change to the
"pcre.recursion_limit" setting (and fix whatever PageTriage's problem is in
some other way), or at least turn it up to something more reasonable than 1024.
I'd expect this is causing problems in other areas of the code, too.

Sam, have you had a chance to look at this yet?

Nope.

We can increase the pcre.recursion_limit again, the value Roan set was overly conservative.

I'm not sure what the availability of various Ops staff is going to be to push this through

(In reply to comment #13)

The regexp recursion limit aside, is using a regexp to check for UTF-8
appropriate? Why not use mb_check_encoding() if available? Other operations in
Language.php do make use of the mb_* functions...

http://php.net/manual/en/function.mb-check-encoding.php

I'm guessing legacy reasons, and the code was never updated.

Please feel free to submit a patch to BZ or a commit to Gerrit. Or is just changing:

		$isutf8 = preg_match( '/^([\x00-\x7f]|[\xc0-\xdf][\x80-\xbf]|' .
				'[\xe0-\xef][\x80-\xbf]{2}|[\xf0-\xf7][\x80-\xbf]{3})+$/', $s );
		if ( $isutf8 ) {
			return $s;
		}

to

		if ( mb_check_encoding( $s, 'UTF-8' )  ) {
			return $s;
		}

enough?

Language::checkTitleEncoding has no unit tests written for it either.

For anyone not following along... Lupo created https://gerrit.wikimedia.org/r/#/c/9126/ and added some unit tests

Code is now live on Wikimedia wikis. Both the regex fix, and the addition of mb_check_encoding() are there. Should be quite a bit quicker. Shouldn't need the recursion limit increasing again either

(In reply to comment #18)

Shouldn't need the recursion limit increasing again either

Are you sure that this low recursion limit isn't breaking things elsewhere in the code too?

(In reply to comment #19)

(In reply to comment #18)

Shouldn't need the recursion limit increasing again either

Are you sure that this low recursion limit isn't breaking things elsewhere in
the code too?

Nope. But leaving it low will bring them up, and maybe we can fix those aswell? ;)

Related URL: https://gerrit.wikimedia.org/r/58306 (Gerrit Change I27203c767d1d3f2f0999b1b1d8a06e8cf68c19ed)