This is similar to the old parser: when discarding a node, actually
evaluate it if short-circuit is not allowed.
Add a whole lot of tests for all possible exceptions.
Move the logic to extract a message from an AFPUserVisibleException away
from the parser, to keep unit tests working.
Bug: T232498
Change-Id: I31ee4e255c6a87dd693b9bcd582539fdf57acd45
This is because there are many filters using this feature. Moreover, it
could make it a little easier to add new arguments, just like dangling
commas in PHP arrays do.
Also re-align the CachingParser code of doLevelFunctions to the one in
the old Parser.
Bug: T153251
Change-Id: Ie4325159f47310788da57415a5e36e62aa4efad0
This will help mitigating problems like T230256 by enforcing that the
requested variables must exist. For now, it will only log bad usages,
thus providing a way to identify affected filters and fix them.
Bug: T230256
Change-Id: I7a61916576e444a56f0e07da7b6e5033346226bd
Using `new LanguageEn()` involved a global, so use a MockObject instead.
Also fix LoggerFactory usage in Tokenizer to use DI instead.
Change-Id: I94d03f9459ab6444e239386eb96a0c2434bfe3dc
IMHO these can be considered unit tests; they were already fast, but now
they're executed in an instant.
This requires several changes: 1 - delay retrieving messages in
AFPUserVisibleException, to avoid having to deal with i18n whenever we
want to test exceptions; 2 - Use some DI for Parser and Tokenizer.
Equivset-dependend tests are also moved to a new class, thus helping to
fix the AF part of T189560.
Change-Id: If4585bf9bb696857005cf40a0d6985c36ac7e7a8
While this is not as important as throwing for too few parameters, IMHO
it's still important to fail in this case. Mostly because if a function
receives too many parameters, chances are that who wrote the filter
didn't do that intendedly, and thus there may be a hidden bug.
Bonus: fix a few docblocks.
Bug: T230803
Change-Id: Iac2931f17b50ace8c8f4c2faa44b3f54ca134c54
This allows a little bit more of abstraction: we can store other data in the
tree, without having to store it in a specific node (e.g. the variables map,
which is still unused). It also adds a few typehints, and specializes
the return value of eval'ing the AST: previously, it was the one of
evalNode, which wasn't guaranteed to be an AFPData. Now we have this
guarantee. Last but not least, we can now measure runtime metrics for
evalTree, which doesn't recurse.
Bonus: fix a check in the old parser, which used the wrong variable when
reporting outofbounds errors.
Change-Id: Iff806793b1d968e9bb6220f1459f3d0ac587c7da
And fix a couple of minor bugs.
Bug: T156096
Depends-On: I3b85087677607573f4fa68681735dc35348dcd87
Change-Id: Ia4c713a1d45827f6a8bc5566a8d8835c49f8108a
Ensure that the variable isn't set before marking it as DUNDEFINED:
that's only for when we cannot use a default, but if the variable is set
we already have one. Most notably, this fixes conditionals handling: right
now, if you have a conditional with an assignment in both
branches, the variable will be undefined. That's obviously wrong, so
it's fixed in this patch.
Plus: catch only AFPExceptions in a test to avoid unintentionally
catching the assert exception; simplify some assignments using wfSetVar.
Depends-On: I446a307e5395ea8cc8ec5ca5d5390b074bea2f24
Change-Id: I8e7f7710b8cb37ada8531b631456a3ce7b27ee45
This patch includes various fixes to how func arguments are handled in
CachingParser:
- Add a comment about a future improvement of checkSyntax, which we
could limit to try building the AST.
- Having enough args for each function is now also checked when
building the AST. This allows implementing the previous point without
stopping to report notenoughargs at syntaxcheck-time (otherwise it'd be
a runtime error). And it also ensure that we check for the params count
inside skipped branches, e.g. inside if/else: these were already only
discovered at runtime in CachingParser. The old parser is not affected
by this change, because when checking syntax it will always execute
all branches, and at runtime it will skip braces altogether.
- Fix arg count for CachingParser, which previously added a bogus param
in case of a function called without parameters. This was fixed for
the other parser in I484fe2994292970276150d2e417801453339e540, and I
just ported the updated fix. Also note that the CachingParser was
already failing for e.g. `count()`, but instead of complaining about
missing arguments, it failed hard when trying to pass NULL to
evalNode.
- Fixed some tests not to use setExpectedException, which caused the
previous point to remain unnoticed: calling that method prevents the
loop from continuing, and thus only the AbuseFilterParser part was
being executed. The new implementation checks the exception ID and is
thus more future-proof if the i18n message changes.
- Fixed some function names in error reporting for the old parser.
- The arg count is now checked outside of the function handlers, thus
it's no more necessary to call checkEnoughArguments at the beginning
of each handler. This also produces clearer error messages in case of
aliases (e.g. set/set_var).
- Check the args count even if some of the args are DUNDEFINED. This is
much easier now that the check is outside of the handler. This will
make syntax check fail for e.g. `contains_any(added_lines)`.
Bug: T156095
Change-Id: I446a307e5395ea8cc8ec5ca5d5390b074bea2f24
Just like we do for functions, it doesn't really make sense to have
keywords separately, in AFPData.
Change-Id: I208a9b1ce2bd12038e9fbcc515c48d604ec80eb8
This patch keeps the current behaviour for everything (since DUNDEFINED
was always casted to boolean false), but handles the cast at a higher
level instead of relying on what AFPData::castTypes will do. This way
it's easier to spot places where we may get DUNDEFINED, and decide how
to handle them one by one.
Change-Id: I1070e15ea03c7dd4a4231b87afbc42240a558581
This is more complicated than the := operator, because the var name
could be a complicated expression, and we have to handle a function
call. This patch only covers the case where the variable name is a
literal, which is enough for WMF production.
Bug: T214674
Change-Id: I6c0f8e95663919a0235b5ccf0c88ad0a539315a7
Follow-up I7831f3ed9f7c0656e0e8f77ded049c20eca682ba, really avoid the leak. My addition was pointless because we need DUNDEFINED, not DEMPTY, and I spent way too much time trying to understand what was still wrong.
Still have to get used to these new names...
Change-Id: I332967f6fb00b67fd355547b19638c95ffa5bba7
As shown in the coverage reports [0], some empty operand logging lines are covered, but no test should have empty operands. I see one of the cause is skipOverBraces keeping $result as is, even if DEMPTY, so turn it into a DUNDEFINED.
[0] - https://doc.wikimedia.org/cover-extensions/AbuseFilter/includes/parser/AbuseFilterParser.php.html
Change-Id: I7831f3ed9f7c0656e0e8f77ded049c20eca682ba
I5ec4ab44c4e88aaf18c0d7b73355d27050beeda7 almost fixed this bug, but we
also have to make it possible to access builtin variables as arrays.
This will only make sense for a few variables (e.g. added_lines and
removed_lines), but I don't think we should validate it when checking
syntax.
Bug: T198531
Change-Id: I417e1b8d4802bbfccd091ce5c7617659cfd1e4ea
Start deprecating "empty" logic operators, and now that we have DEMPTY, simplify handling of empty function arguments introduced in Ica3e49f5b00595a95513d9683732e490aa7aae17.
Bug: T156096
Change-Id: Ied6b385e8690b6cc6e69afcf614389f737ab95bd
This patch includes:
* Making it possible to access offsets of a DNONE (returning a DNONE)
* Initializing user-defined variables as DNONE inside short-circuited branches
* Make DNONE propagate with other operators
* Make DNONE count as false for logic operators
* Remove a now-outaded bit in doLevelAtom. In case of shortcircuit,
$result is now DNONE instead of DNULL, and thus it's possible to
access offsets of it. Performance++!
* Don't allow modifying or adding an element of a DNONE as if it were an
array (to avoid inconsistencies)
This re-applies Id85c673337fa90a3782fd22eb9690cd996967111 with several fixes.
NOTE: Haven't tested locally, although I'm pretty confident thanks to
the amount of tests added.
Bug: T214674
Bug: T228677
Change-Id: I5ec4ab44c4e88aaf18c0d7b73355d27050beeda7
Currently we strongly abuse (pardon the pun) the AbuseFilter class: its
purpose should be to hold static functions intended as generic utility
functions (e.g. to format messages, determine whether a filter is global
etc.), but we actually use it for all methods related to running filters.
This patch creates a new class, AbuseFilterRunner, containing all such
methods, which have been made non-static. This leads to several
improvements (also for related methods and the parser), and opens the
way to further improve the code.
Aside from making the code prettier, less global and easier to test,
this patch could also produce a performance improvement, although I
don't have tools to measure that.
Also note that many public methods have been removed, and almost any of
them has been made protected; a couple of them (the ones used from outside)
are left for back-compat, and will be removed in the future.
Change-Id: I2eab2e50356eeb5224446ee2d0df9c787ae95b80
This should allow more flexibility when checking syntax, and a saner
behaviour overall.
Aside from not throwing exception in certain cases, the results should
be almost equal to the ones you would get without this patch. However,
there are still a few things to improve (which for convenience I wrote
inside the parser test) and many to test.
Bug: T204654
Depends-On: I69bfec45c76509fb1112641393f78e8d8834adcd
Change-Id: I5a14d4b2bc3ffd9caaaa095f16f36b9b6009db05
$data and $type are meant to be read-only and should have getter
functions, but as usual they're just public. Add getter methods, a
comment with a @private annotation and remove usages in our codebase.
Change-Id: I5e51efc9f982a4e340b48d20cb1b38a75bb10021
This property is meant to be private, since it has all kinds of
getters/setters, aside from one which is introduced in this patch.
Change-Id: I217b1e22cabd3c0468c84b1d6a69a6ed3c6fa8e6
Split from I5a14d4b2bc3ffd9caaaa095f16f36b9b6009db05.
This adds a new data type to use for empty AFPDatas. Using NULL for that
makes it impossible to distinguish cases where we really got a null
value, and cases where there was nothing to parse.
For now, DNONE is the same as DNULL, but I've explicited DNULL where
necessary. A subsequent patch will make proper use of DNONE.
Bug: T156096
Change-Id: I69bfec45c76509fb1112641393f78e8d8834adcd
As explained on phabricator, they don't work with shortcircuit, so they
already fail for all filters using them. Plus IMHO it's an unnecessary
deviation from PHP's behaviour, given that this syntax doesn't do what
users may expect.
Bug: T218906
Change-Id: If9e7545e14044c8dc3b4163bb6fca8ab0683b9fa
Instead of relying on static methods and members in the AbuseFilter
class, move everything related to conditions inside the Parser, as the
amount of used conditions is something pertaining a single
AbuseFilter(Caching)Parser instance.
This change requires changing some signatures and adding parameters,
but will make introducing the new AbuseFilterRunner class easier (and
that will clean signatures, too).
Depends-On: I5b29ff556eca45fe59d15e2e3df4d06f1f6b3934
Change-Id: I7c1ea17adf7f42cf9260d416906bfbf3b8a20688
