For fixing bugs like T65632, T105325, or T264104, we will need
to update code in more than one place at once. To prevent
regressions, create an integration test which tests the whole
pipeline, from the request submission to variable evaluation.
Edits are simulated using action=edit API call because the hook
AbuseFilter uses is run from EditPage.
To increase confidence in test coverage, remove some annotations
from AbuseFilterConsequencesTest or make them less greedy.
Ideally, it would only test consequences.
This patch includes refactoring of AbuseFilterCreateAccountTestTrait
which now only inserts the user into the database if it really
should be created.
It also restores test coverage of some other classes.
Change-Id: I661f4e0e2bcac4770e499708fca4e4e153f31fed
Test that null edits do not trigger filters, but sole
content model change does.
Also do some cleanup in AbuseFilterConsequencesTest.
For better isolation, do not access the service
container and do not initialize objects in
the constructor.
Change-Id: I043ecb312226a69d1f485a8382d558ccb899a270
Using Title::newFromText is parsing the string, which is expensive.
Just use Title::makeTitle when the result is known.
editPage() can take a Title or WikiPage instead of a string, avoid
creation of Title there.
The default ns on editPage() is only needed when giving a string
Change-Id: Ie303b9e6d6b8d6ac80286059f8e86bfc76b779af
This was most definitely my intention when I introduced the concept of
"generic vars", so it's a bit surprising to discover, 3.5 years later,
that the timestamp isn't computed there.
Also make the timestamp always be a string for consistency, since that's
the type documented on mw.org. I've manually checked all filters on
Wikimedia wikis using the timestamp variable, and added explicit int
casts where needed (although I think they'd still work due to implicit
casts).
Change-Id: Ib6e15225dd95c2eead7e48c200d203d6918e0c18
Introduce shorter methods, one for each steps, so that it's easier to
understand what the code is doing and figure out if the order makes
sense. The ConsequencesExecutor test is now a proper unit test. Also
simplify AbuseFilterConsequencesTest, removing old/wrong logic and
fixing two expected values that were actually wrong (but worked because
of the aforementioned wrong logic).
The only functional changes should be:
- We pick the longest block *after* checking the ConsequenceDisabler
consequences, so e.g. if a filter has a long block + warn and another
filter has a shorter block, we still keep the second one if warn will
disable the block.
- Remove disallow in presence of dangerous actions after checking
ConsequenceDisabler's and deduplicating blocks. Otherwise we may
remove disallow for filters where block (etc.) doesn't end up being
disabled. We may also want to consider not removing disallow at all,
now that messages are customizable.
Bug: T303059
Change-Id: If00adbf2056758222eaaea70b16d3b4f89502c20
WikiPage::factory() is deprecated since 1.36 and should be replaced
with WikiPageFactory::newFromTitle().
Bug: T297688
Change-Id: I85d3566519ab977aad8c517cc48fc8c271e5589a
Regression tests to make sure T286140 does not
happen again.
In the process, discovered what caused that bug
with afl_rev_id not being set: EditRevUpdater::updateRev()
compares the WikiPage given in the PageSaveComplete hook
to the one given to it by AbuseFilterHooks from
onEditFilterMergedContent, and compares the two using
`===`, meaning that they must refer to the same underlying
object. That bug was caused because AbuseFilterHooks
changed to providing a different object, despite still
referring to the same underlying page.
We should probably change that behavior in EditRevUpdater,
but for now updated AbuseFilterConsequencesTest to pass
the same object around by using RequestContext::setWikiPage()
and providing the WikiPage object to
MediaWikiIntegrationTestCase::editPage().
Bug: T286140
Change-Id: I6562f513c463538af6b59b12a64564b254024613
Avoid strtotime and compare TS_MW timestamps
Set a fake time to get the same block expiry for relative times
Bug: T272236
Depends-On: I1357d3a78538b8bcb2a3507d86f35371e3f26d47
Change-Id: I5447953c5a0d7ecf4534f4ac4bc2260fa9f42117
Mainly constructor and conditions limit, which can be removed from
ConsequencesTest (where it was very slow).
Additionally, inject globals into FilterRunner.
Change-Id: I56ca67de6878dbc2185038faae3eb2b04fb56be9
So everything can be loaded using PSR-4. These classes weren't renamed,
nor the alias for the AbuseFilter class was deprecated, because they
should be refactored first.
Change-Id: Ia328db58eb326968edf5591daac9bacf8c2f75da
This makes VariableHolder a true value object, and introduces a
stateless service, VariableManager, to operate on it.
Note, in theory, this new service is still cyclically coupled with
LazyVariableComputed. However, it's now two stateless service being
coupled, not two smart/god value objects, so we've still earned
something. For now, the dependency is hidden by using a callback. Some
alternatives for that are mentioned in a code comment.
Bug: T261069
Change-Id: I2f2c84c8e91472ba36084a8bbb4a923f6e04354b
Skip a test that fails with
Wikimedia\Rdbms\DBQueryError: Error 5: database is locked
Function: Wikimedia\Rdbms\Database::beginIfImplied (MediaWiki\Extension\AbuseFilter\FilterLookup::getAllActiveFiltersInGroupFromDB)
Probably due to some concurrency issue caused by the duplicate connection, and also with
Wikimedia\Rdbms\DBQueryError: Error 1: no such table: unittest_external_abuse_filter
Function: MediaWiki\Extension\AbuseFilter\FilterLookup::getAllActiveFiltersInGroupFromDB
for unknown reasons.
Move the mwGlobals override inside the test to avoid the same "database is locked" error
on every other test in that class.
Bug: T251967
Change-Id: I552a8d1fa532941f630fd734e590993e7462aeb0
This is the last big step towards moving Consequences-related things away from
AbuseFilterRunner. There's still some cleanup to do (+ write proper tests), but
this should really be the last important code change.
Change-Id: I347795fe93ba496c43b1d5cfc9ba6e1326842c06
Add a script to migrate the columns (which can also
be executed in dry run), and a config option with the migration stage
(defaults to SCHEMA_COMPAT_OLD).
Some of the script-related code is stolen from
Ic755526d5f989c4a66b1d37527cda235f61cb437.
Bug: T220791
Change-Id: I7460a2d63f60c2933b36f8383a8abdbba8649e12
The consequence-taking logic is moved away from AbuseFilterRunner, to
dedicated classes. There's now one class per consequence, encapsulating
everything it needs to take the consequence.
Several interfaces allow customizing different types of consequences.
Every "special check" in AbuseFilter was generalized to use these
interfaces, rather than knowing how to handle each consequence.
Adding more consequences from other extensions will also be easier, and
it should happen via a hook (not a global), returning a class that
implements Consequence. The BCConsequence class was temporarily added
for legacy custom consequences.
A ConsequenceFactory class is added to instantiate consequences; this
would possibly benefit from using ObjectFactory, but it doesn't because
it would also reduce readability (although we might do that in the
future).
These classes are still not covered by unit tests, and this is left to
do for later. The new unit tests should mostly replace
AbuseFilterConsequencesTest. @covers tag were added to keep the status
quo (i.e. code that was considered covered while in AbuseFilterRunner
will still be considered covered), although we'll have to adjust them.
Change-Id: Ia1a9a8bbf55ddd875dfd5bbc55fcd612cff568ef
This way we don't have special cases in executeFilterActions, and instead, we execute
all actions in the same place. In turn, this is going to ease the
transition to a new consequences system: next step is refactoring this
code into a service with proper DI etc.
Bug: T204447
Change-Id: I8134ecc41fbecdbed99faf406e9e3ca91b6123b9
This commit removes several tests from AbuseFilterConsequences, thus
speeding it up a lot (especially because these tests were very slow,
with each test *case* taking up to 30s in the coverage job).
Everything is now covered by the new AbuseFilterFilterProfilerTest
which, although not being a pure unit test, is much much faster than
*Consequences.
Change-Id: Ic6b16d23ec99abee287f36093b8573505f9c613a
This service is responsible for the blockautopromote feature:
(un)block autopromotion and check status.
The patch mostly moves code from static methods to the new class
and relaxes type hints (e.g. from User to UserIdentity).
Change-Id: I79a72377881cf06717931cd09af12f3b8e5f3e3f
So that sysadmins can further customize the extension. It was also wrong
to use the same variable for many different things.
Note that there's no associated patch in wmf-config because we use the
defaults. However, before merging this patch, please recheck that
AbuseFilterRestrictions and AbuseFilterDisallowGlobalLocalBlocks aren't
used there (https://codesearch.wmflabs.org/operations/?q=AbuseFilterDisallowGlobalLocalBlocks%7CAbuseFilterRestrictions&i=nope&files=&repos=)
Bug: T175221
Change-Id: I7581b3ee6d9d11a6cf1599b8ff874e8c3d54adf4
The logic about action IDs and the persistent buffer is now encapsulated
inside a single service, which is a step towards getting rid of global
state in the AbuseFilter class, and reducing the responsibilities of the
Runner.
An important change made here is that we now require a LinkTarget rather
than a Title. This removes a dependency on the Title class (a monster
object), makes tests simpler, and denies the need to inject a
TitleFactory. This means living without some bits of context (e.g. we're
no longer using makeTitleSafe to ensure a valid title, and we have to
build a "prefixedtext" manually), but this shouldn't be a problem, given
that the titles are only used to create a cache key: invalid titles are
not a problem, and concatenating namespace + title should always be
sufficient.
Bug: T265370
Change-Id: Iff59cd3d889454a482a89c16691bfefcc5ec0a12
