Pattern syntax
Getting started with rule writing? Try the Semgrep Tutorial 🎓
This document describes Semgrep’s pattern syntax. You can also see pattern examples by language. In the command line, patterns are specified with the flag --pattern (or -e). Multiple
coordinating patterns may be specified in a configuration file. See
rule syntax for more information.
Pattern matching
Pattern matching searches code for a given pattern. For example, the
expression pattern 1 + func(42) can match a full expression or be
part of a subexpression:
foo(1 + func(42)) + bar()
In the same way, the statement pattern return 42 can match a top
statement in a function or any nested statement:
def foo(x):
if x > 1:
if x > 2:
return 42
return 42
Ellipsis operator
The ... ellipsis operator abstracts away a sequence of zero or more
items such as arguments, statements, parameters, fields, characters.
The ... ellipsis can also match any single item that is not part of
a sequence when the context allows it.
See the use cases in the subsections below.
Function calls
Use the ellipsis operator to search for function calls or
function calls with specific arguments. For example, the pattern insecure_function(...) finds calls regardless of its arguments.
insecure_function("MALICIOUS_STRING", arg1, arg2)
Functions and classes can be referenced by their fully qualified name, e.g.,
django.utils.safestring.mark_safe(...)ormark_safe(...)System.out.println(...)orprintln(...)
You can also search for calls with arguments after a match. The pattern func(1, ...) will match both:
func(1, "extra stuff", False)
func(1) # Matches no arguments as well
Or find calls with arguments before a match with func(..., 1):
func("extra stuff", False, 1)
func(1) # Matches no arguments as well
The pattern requests.get(..., verify=False, ...) finds calls where an argument appears anywhere:
requests.get(verify=False, url=URL)
requests.get(URL, verify=False, timeout=3)
requests.get(URL, verify=False)
Match the keyword argument value with the pattern $FUNC(..., $KEY=$VALUE, ...).
Method calls
The ellipsis operator can also be used to search for method calls.
For example, the pattern $OBJECT.extractall(...) matches:
tarball.extractall('/path/to/directory') # Oops, potential arbitrary file overwrite
You can also use the ellipsis in chains of method calls. For example,
the pattern $O.foo(). ... .bar() will match:
obj = MakeObject()
obj.foo().other_method(1,2).again(3,4).bar()
Function definitions
The ellipsis operator can be used in function parameter lists or in the function body. To find function definitions with mutable default arguments:
pattern: |
def $FUNC(..., $ARG={}, ...):
...
def parse_data(parser, data={}): # Oops, mutable default arguments
pass
The YAML | operator allows for multiline strings.
The ellipsis operator can match the function name.
Match any function definition:
Regular functions, methods, and also anonymous functions (such as lambdas).
To match named or anonymous functions use an ellipsis ... in place of the name of the function.
For example, in JavaScript the pattern function ...($X) { ... } matches
any function with one parameter:
function foo(a) {
return a;
}
var bar = function (a) {
return a;
};
Class definitions
The ellipsis operator can be used in class definitions. To find classes that inherit from a certain parent:
pattern: |
class $CLASS(InsecureBaseClass):
...
class DataRetriever(InsecureBaseClass):
def __init__(self):
pass
The YAML | operator allows for multiline strings.
Ellipsis operator scope
The ... ellipsis operator matches everything in its current scope. The current scope of this operator is defined by the patterns that precede ... in a rule. See the following example:
Semgrep matches the first occurrence of bar and baz in the test code as these objects fall under the scope of foo and .... The ellipsis operator does not match the second occurrence of bar and baz as they are not inside of the function definition, therefore these objects in their second occurrence are not inside the scope of the ellipsis operator.
Strings
The ellipsis operator can be used to search for strings containing any data. The pattern crypto.set_secret_key("...") matches:
crypto.set_secret_key("HARDCODED SECRET")
This also works with constant propagation.
In languages where regular expressions use a special syntax
(for example JavaScript), the pattern /.../ will match
any regular expression construct:
re1 = /foo|bar/;
re2 = /a.*b/;
Binary operations
The ellipsis operator can match any number of arguments to binary operations. The pattern $X = 1 + 2 + ... matches:
foo = 1 + 2 + 3 + 4
Containers
The ellipsis operator can match inside container data structures like lists, arrays, and key-value stores.
The pattern user_list = [..., 10] matches:
user_list = [8, 9, 10]
The pattern user_dict = {...} matches:
user_dict = {'username': 'password'}
The pattern user_dict = {..., $KEY: $VALUE, ...} matches the following and allows for further metavariable queries:
user_dict = {'username': 'password', 'address': 'zipcode'}
You can also match just a key-value pair in
a container, for example in JSON the pattern "foo": $X matches
just a single line in:
{ "bar": True,
"name": "self",
"foo": 42
}
Conditionals and loops
The ellipsis operator can be used inside conditionals or loops. The pattern:
pattern: |
if $CONDITION:
...
The YAML | operator allows for multiline strings.
matches:
if can_make_request:
check_status()
make_request()
return
A metavariable can match a conditional or loop body if the body statement information is re-used later. The pattern:
pattern: |
if $CONDITION:
$BODY
matches:
if can_make_request:
single_request_statement()
Half or partial statements can't be matches; both of the examples above must specify the contents of the condition’s body (e.g., $BODY or ...), otherwise they are not valid patterns.
Matching single items with an ellipsis
Ellipsis ... is generally used to match sequences of similar elements.
However, you can also match single item using ellipsis ... operator.
The following pattern is valid in languages with a C-like
syntax even though ... matches a single Boolean value rather
than a sequence:
if (...)
return 42;
Another example where a single expression is matched by an ellipsis is the right-hand side of assignments:
foo = ...;
However, matching a sequence of items remains the default meaning of an
ellipsis. For example, the pattern bar(...) matches bar(a),
but also bar(a, b) and bar(). To force a match on a single item,
use a metavariable as in bar($X).
Metavariables
Metavariables are an abstraction to match code when you don’t know the value or contents ahead of time, similar to capture groups in regular expressions.
Metavariables can be used to track values across a specific code scope. This includes variables, functions, arguments, classes, object methods, imports, exceptions, and more.
Metavariables look like $X, $WIDGET, or $USERS_2. They begin with a $ and can only
contain uppercase characters, _, or digits. Names like $x or $some_value are invalid.
Expression metavariables
The pattern $X + $Y matches the following code examples:
foo() + bar()
current + total
Import metavariables
Metavariables can also be used to match imports. For example, import $X matches:
import random
Reoccurring metavariables
Re-using metavariables shows their true power. Detect useless assignments:
pattern: |
$X = $Y
$X = $Z
Useless assignment detected:
initial_value = 10 # Oops, useless assignment
initial_value = get_initial_value()
The YAML | operator allows for multiline strings.
Literal Metavariables
You can use "$X" to match any string literal. This is similar
to using "...", but the content of the string is stored in the
metavariable $X, which can then be used in a message
or in a metavariable-regex.
You can also use /$X/ and :$X to respectively match
any regular expressions or atoms (in languages that support
those constructs, e.g., Ruby).
Because literal metavariables bind to strings that may not be valid code, if you want to match them in more detail with a metavariable-pattern, you must specify generic language inside the metavariable-pattern. For example:
rules:
- id: match-literal-string
languages:
- python
severity: INFO
message: Found "$STRING"
patterns:
- pattern: '"$STRING"'
- metavariable-pattern:
language: generic
metavariable: $STRING
pattern: "literal string contents"
Typed metavariables
Syntax
Typed metavariables only match a metavariable if it’s declared as a specific type.
Java:
For example, to look for calls to the log method on Logger objects.
A simple pattern for this purpose could use a metavariable for the Logger object.
pattern: $LOGGER.log(...)
But if we are concerned about finding calls to the Math.log() method as well, we can use a typed metavariable to put a type constraint on the $LOGGER metavariable.
pattern: (java.util.logging.Logger $LOGGER).log(...)
Alternatively, if we want to capture more logger types, for example custom logger types, we could instead add a constraint to the type of the argument in this method call instead.
pattern: $LOGGER.log(java.util.logging.LogRecord $RECORD)
C:
In this example in C, we want to capture all cases where something is compared to a char array. We start with a simple pattern that looks for comparison between two variables.
pattern: $X == $Y
We can then put a type constraint on one of the metavariables used in this pattern by turning it into a typed metavariable.
pattern: $X == (char *$Y)
int main() {
char *a = "Hello";
int b = 1;
// Matched
if (a == "world") {
return 1;
}
// Not matched
if (b == 2) {
return -1;
}
return 0;
}
Go:
The syntax for a typed metavariable in Go looks different from the syntax for Java.
In this Go example we look for calls to the Open function, but only on an object of the zip.Reader type.
pattern: |
($READER : *zip.Reader).Open($INPUT)
func read_file(reader *zip.Reader, filename) {
// Matched
reader.Open(filename)
dir := http.Dir("/")
// Not matched
f, err := dir.Open(c.Param("file"))
}
For Go, Semgrep currently does not recognize the type of all variables that are declared on the same line. That is, the following will not take both a and b as ints: var a, b = 1, 2
TypeScript:
In this example, we want to look for uses of the DomSanitizer function.
pattern: ($X: DomSanitizer).sanitize(...)
constructor(
private _activatedRoute: ActivatedRoute,
private sanitizer: DomSanitizer,
) { }
ngOnInit() {
// Not matched
this.sanitizer.bypassSecurityTrustHtml(DOMPurify.sanitize(this._activatedRoute.snapshot.queryParams['q']))
// Matched
this.sanitizer.bypassSecurityTrustHtml(this.sanitizer.sanitize(this._activatedRoute.snapshot.queryParams['q']))
}
Using typed metavariables
Type inference applies to the entire file! One common way to use typed metavariables is to check for a function called on a specific type of object. For example, let's say you're looking for calls to a potentially unsafe logger in a class like this:
class Test {
static Logger logger;
public static void run_test(String input, int num) {
logger.log("Running a test with " + input);
test(input, Math.log(num));
}
}
If you searched for $X.log(...), you can also match Math.log(num). Instead, you can search for (Logger $X).log(...) which gives you the call to logger. See the rule logger_search.
Since matching happens within a single file, this is only guaranteed to work for local variables and arguments. Additionally, Semgrep currently understands types on a shallow level. For example, if you have int[] A, it will not recognize A[0] as an integer. If you have a class with fields, you will not be able to use typechecking on field accesses, and it will not recognize the class’s field as the expected type. Literal types are understood to a limited extent. Expanded type support is under active development.
Ellipsis metavariables
You can combine ellipses and metavariables to match a sequence
of arguments and store the matched sequence in a metavariable.
For example the pattern foo($...ARGS, 3, $...ARGS) will
match:
foo(1,2,3,1,2)
When referencing an ellipsis metavariable in a rule message or metavariable-pattern, include the ellipsis:
- message: Call to foo($...ARGS)
Anonymous metavariables
Anonymous metavariables are used to specify that a metavariable exists in the pattern you want to capture.
An anonymous metavariable always takes the form $_. Variables such as $_1 or $_2 are not anonymous. You can use more than one anonymous metavariable in a rule definition.
For example, if you want to specify that a function should always have 3 arguments, then you can use anonymous metavariables:
- pattern: def function($_, $_, $_)
An anonymous metavariable does not produce any binding to the code it matched. This means it does not enforce that it matches the same code at each place it is used. The pattern:
- pattern: def function($A, $B, $C)
is not equivalent to the former example, as $A, $B, and $C bind to the code that matched the pattern. You can then use $A or any other metavariable in your rule definition to specify that specific code. Anonymous metavariables cannot be used this way.
Anonymous metavariables also communicate to the reader that their values are not relevant, but rather their occurrence in the pattern.
Metavariable unification
For search mode rules, metavariables with the same name are treated as the same metavariable within the patterns operator. This is called metavariable unification.
For taint mode rules, patterns defined within pattern-sinks and pattern-sources still unify. However, metavariable unification between pattern-sinks and pattern-sources is not enabled by default.
To enforce unification, set taint_unify_mvars: true under the rule options key. When taint_unify_mvars: true is set, a metavariable defined in pattern-sinks and pattern-sources with the same name is treated as the same metavariable. See Metavariables, rule message, and unification for more information.
Display matched metavariables in rule messages
Display values of matched metavariables in rule messages. Add a metavariable to the rule message (for example Found $X) and Semgrep replaces it with the value of the detected metavariable.
To display matched metavariable in a rule message, add the same metavariable as you are searching for in your rule to the rule message.
- Find the metavariable used in the Semgrep rule. See the following example of a part Semgrep rule (formula):
This formula uses
- pattern: $MODEL.set_password(…)$MODELas a metavariable. - Insert the metavariable to rule message:
- message: Setting a password on $MODEL - Use the formula displayed above against the following code:
user.set_password(new_password)
The resulting message is:
Setting a password on user
Run the following example in Semgrep Playground to see the message (click Open in Editor, and then Run, unroll the 1 Match to see the message):
If you're using Semgrep's advanced dataflow features, see documentation of experimental feature Displaying propagated value of metavariable.
Equivalences
Semgrep automatically searches for code that is semantically equivalent.
Imports
Equivalent imports using aliasing or submodules are matched.
The pattern subprocess.Popen(...) matches:
import subprocess.Popen as sub_popen
sub_popen('ls')
The pattern foo.bar.baz.qux(...) matches:
from foo.bar import baz
baz.qux()
Constants
Semgrep performs constant propagation.
The pattern set_password("password") matches:
HARDCODED_PASSWORD = "password"
def update_system():
set_password(HARDCODED_PASSWORD)
Basic constant propagation support like in the example above is a stable feature. Experimentally, Semgrep also supports intra-procedural flow-sensitive constant propagation.
The pattern set_password("...") also matches:
def update_system():
if cond():
password = "abc"
else:
password = "123"
set_password(password)
It is possible to disable constant propagation in a per-rule basis via the options rule field.
Associative and commutative operators
Semgrep performs associative-commutative (AC) matching. For example, ... && B && C will match both B && C and (A && B) && C (i.e., && is associative). Also, A | B | C will match A | B | C, and B | C | A, and C | B | A, and any other permutation (i.e., | is associative and commutative).
Under AC-matching metavariables behave similarly to .... For example, A | $X can match A | B | C in four different ways ($X can bind to B, or C, or B | C). In order to avoid a combinatorial explosion, Semgrep will only perform AC-matching with metavariables if the number of potential matches is small, otherwise it will produce just one match (if possible) where each metavariable is bound to a single operand.
Using options it is possible to entirely disable AC-matching. It is also possible to treat Boolean AND and OR operators (e.g., && in || in C-family languages) as commutative, which can be useful despite not being semantically accurate.
Deep expression operator
Use the deep expression operator <... [your_pattern] ...> to match an expression that could be deeply nested within another expression. An example is looking for a pattern anywhere within an if statement. The deep expression operator matches your pattern in the current expression context and recursively in any subexpressions.
For example, this pattern:
pattern: |
if <... $USER.is_admin() ...>:
...
matches:
if user.authenticated() and user.is_admin() and user.has_group(gid):
[ CONDITIONAL BODY ]
The deep expression operator works in:
ifstatements:if <... $X ...>:- nested calls:
sql.query(<... $X ...>) - operands of a binary expression:
"..." + <... $X ...> - any other expression context
Limitations
Statements types
Semgrep handles some statement types differently than others, particularly when searching for fragments inside statements. For example, the pattern foo will match these statements:
x += foo()
return bar + foo
foo(1, 2)
But foo will not match the following statement (import foo will match it though):
import foo
Statements as expressions
Many programming languages differentiate between expressions and statements. Expressions can appear inside if conditions, in function call arguments, etc. Statements can not appear everywhere; they are sequence of operations (in many languages using ; as a separator/terminator) or special control flow constructs (if, while, etc.).
foo() is an expression (in most languages).
foo(); is a statement (in most languages).
If your search pattern is a statement, Semgrep will automatically try to search for it as both an expression and a statement.
When you write the expression foo() in a pattern, Semgrep will visit every expression and sub-expression in your program and try to find a match.
Many programmers don't really see the difference between foo() and foo();. This is why when one looks for foo(); Semgrep thinks the user wants to match statements like a = foo();, or print(foo());.
Note that in some programming languages such as Python, which does not use semicolons as a separator or terminator, the difference between expressions and statements is even more confusing. Indentation in Python matters, and a newline after foo() is really the same than foo(); in other programming languages such as C.
Partial expressions
Partial expressions are not valid patterns. For example, the following is invalid:
pattern: 1+
A complete expression is needed (like 1 + $X)
Ellipses and statement blocks
The ellipsis operator does not jump from inner to outer statement blocks.
For example, this pattern:
foo()
...
bar()
matches:
foo()
baz()
bar()
and also matches:
foo()
baz()
if cond:
bar()
but it does not match:
if cond:
foo()
baz()
bar()
because ... cannot jump from the inner block where foo() is, to the outer block where bar() is.
Partial statements
Partial statements are partially supported. For example,
you can just match the header of a conditional with if ($E),
or just the try part of an exception statement with try { ... }.
This is especially useful when used in a pattern-inside to restrict the context in which to search for other things.
Other partial constructs
It is possible to just match the header of a function (without its body),
for example int foo(...) to match just the header part of the
function foo. In the same way, you can just match a class header
(e.g., with class $A).
Deprecated features
String matching
String matching has been deprecated. You should use metavariable-regex instead.
Search string literals within code with Perl Compatible Regular Expressions (PCRE).
The pattern requests.get("=~/dev\./i") matches:
requests.get("api.dev.corp.com") # Oops, development API left in
To search for specific strings, use the syntax "=~/<regexp>/". Advanced regexp features are available, such as case-insensitive regexps with '/i' (e.g., "=~/foo/i"). Matching occurs anywhere in the string unless the regexp ^ anchor character is used: "=~/^foo.*/" checks if a string begins with foo.
Not finding what you need in this doc? Ask questions in our Community Slack group, or see Support for other ways to get help.