This filter is fully functional. The old filter is still enabled by
default for a transitional period in case the new one suddenly has
issues.
Change-Id: I4aea5f00c62420108030e60e79d5bf34e913e95d
I am pretty sure all of the behavior documented in these tests is a bad
idea. It is possible that we can fix it since some of those features
are probably unused, but for now those tests will serve as a
documentation of the current behavior.
Change-Id: Ia2a2f57a538d7aef2ac73fb2e47fe82dd5d5e09a
substr_count() is just as fast as looped strpos() when there are no
matches, and gets faster as the number of matches increases.
Note that this introduces a small change in behavior when the needle
is composed of repeated substrings, e.g. 'asdasdasd' or 'aa', and
haystack is such that the needle can be matched in overlapping
positions, e.g. 'asdasdasdasd' or 'aaaaa'. The old implementation
counted overlapping matches, the new one doesn't. I don't think this
behavior was intentional and I don't think this change will cause any
real problems.
Change-Id: Icc905ca34bf08d63e969787a5e3c119d498bf878
Previously, 'false & a == b' would actually execute the comparison and
count it against the condition limit, while 'false & (a == b)' wouldn't.
They behave the same now.
mShortCircuit was only checked for the most potentially expensive
operations (computing functions and getting variables), all the other
operations on bogus values generated by this would be executed and the
results ignored later.
This probably doesn't noticeably improve performance, but it corrects
how the condition limit is counted.
Bug: T43693
Change-Id: Id1d5f577b14b6ae6d987ded12689788eb7922474
With the new behavior, the number of conditions in incremented when:
* Evaluating a function
* Evaluating a comparison operator (== === != !== < > <= >= =)
* Evaluating a keyword (in like matches contains rlike irlike regex)
Previously, the number of conditions was incremented when:
* Evaluating a function
* Entering the comparison operator evaluation mode
This resulted in a number of surprising behaviors. In particular:
* '(((a == b)))' counted as 4 conditions, not 1
* 'contains_any(a, b, c)' counted as 5 conditions, not 1
* 'a == b == c' counted as 1 condition, not 2
* 'a in b + c in d + e in f' counted as 1 condition, not 3
* 'true' counted as 1 condition, not 0
It is still possible to easily cheat the count by rewriting comparisons
as arithmetic operations. I believe this is meant to advise users of
the complexity of their rules and not really enforce strict limits.
Bug: T132190
Change-Id: I897769db4c2ceac802e3ae5d6fa8e9c9926ef246
* Move AbuseFilterParser::nextToken() and the various AbuseFilterParser
properties that accompanied it to a new class, AbuseFilterTokenizer.
* Tokenize rules eagerly and cache the result in APC.
Change-Id: I15f5b5b65e8c4ec4fba3000d7c9fd78b98967d1d
No functional changes.
* Don't include $code as part of the return value; it is ignored anyway.
* Removed AbuseFilterParser::lastHandledToken and AFPParserState::lastInput,
because AbuseFilterParser::nextToken() no longer calls itself recursively.
* The regular expression that matches operators is no longer constructed
dynamically, but hard-coded into the class. To make sure it does not drift
apart from the more legible AbuseFilterParser::$mOps, add a unit test that
constructs the regex dynamically as before and compares it to
AbuseFilterParser::OPERATOR_RE.
* AbuseFilterParser::RADIX_RE ditto.
Change-Id: I9c23b60759ed2f4c73a9b480243b16bbce5a208f
If we don't map '\-' and '\+' to themselves, the leading slash gets escaped,
and the resultant pattern only matches a literal slash.
Bug: 67670
Change-Id: Ifa1e3edd6f41985a3bb97bfb1497985f8fa64af5
I've also added myself to the credits file as I'm the only
maintainer of this extension for a while now.
Change-Id: Id998172ea2abd70b8243de9db1a96cc2cfa47a64
The abusefilter array test failed because length( ['a', 'b', 'c'] )
returned 12 instead of 6. That was du to it converted the array
to a string with new line seperated values first before measuring
the string length. Changed that behaviour to act like the php count()
function or the python len() function which seems far more useful to me.
The old behaviour can be established using length( string( array ) ).
Change-Id: I16646891837c9743ca5af2dd328077a7225bb5f1
ATTENTION! This may break filters that rely on "added_lines contains 'bla-bla'" syntax. They'll need to be replaced with "string(added_lines) contains 'bla-bla'"
* Introduce := operator for setting variables
* Throw an exception when user tries to override built-in variable
* Fix UTF-8 handling in fnmatch() fallback
* Copy three main abuse filters from enwiki to test suite
* Fix update.php integration
* Use strcspn to scan ahead for long regions of uninteresting text in string handling (performance).
* Remove cruft specific to my system in phpTest.php.
* Remove a test that was in incorrect syntax, and useless without adding variable support.
I've made it more performant and fixed a few bugs by using regexes
instead of PHP loops, where possible, under the assumption that the
PCRE parser is more efficient than the same thing implemented in pure PHP.
Also, I'm now passing the same string around and calculating offsets, which
Tim tells me is far more performant than continually truncating the same string.
All tests still pass, with the exception of string.t, which I've modified
to remove the offending code, which never worked.
