14.12 Sandboxed Evaluation

The racket/sandbox module provides utilities for creating “sandboxed” evaluators, which are configured in a particular way and can have restricted resources (memory and time), filesystem and network access, and much more. Sandboxed evaluators can be configured through numerous parameters — and the defaults are set for the common use case where sandboxes are very limited.

The returned evaluator operates in an isolated and limited environment. In particular, filesystem access is restricted, which may interfere with using modules from the filesystem that are not in a collection. See below for information on the allow-for-require, allow-for-load, and allow-read arguments. When language is a module path or when requires is provided, the indicated modules are implicitly included in the allow-for-require list. (For backward compatibility, non-module-path? path strings are allowed in requires; they are implicitly converted to paths before addition to allow-for-require.)

Each input-program or module-decl argument provides a program in one of the following forms:

• an input port used to read the program;

• a string or a byte string holding the complete input;

• a path that names a file holding the input; or

• an S-expression or a syntax object, which is evaluated as with eval (see also get-uncovered-expressions).

In the first three cases above, the program is read using sandbox-reader, with line-counting enabled for sensible error messages, and with 'program as the source (used for testing coverage). In the last case, the input is expected to be the complete program, and is converted to a syntax object (using 'program as the source), unless it already is a syntax object.

The returned evaluator function accepts additional expressions (each time it is called) in essentially the same form: a string or byte string holding a sequence of expressions, a path for a file holding expressions, an S-expression, or a syntax object. If the evaluator receives an eof value, it is terminated and raises errors thereafter. See also kill-evaluator, which terminates the evaluator without raising an exception.

For make-evaluator, multiple input-programs are effectively concatenated to form a single program. The way that the input-programs are evaluated depends on the language argument:

• The language argument can be a module path (i.e., a datum that matches the grammar for module-path of require).

  In this case, the input-programs are automatically wrapped in a module, and the resulting evaluator works within the resulting module’s namespace.

• The language argument can be a list starting with 'special, which indicates a built-in language with special input configuration. The possible values are '(special r5rs) or a value indicating a teaching language: '(special beginner), '(special beginner-abbr), '(special intermediate), '(special intermediate-lambda), or '(special advanced).

  In this case, the input-programs are automatically wrapped in a module, and the resulting evaluator works within the resulting module’s namespace. In addition, certain parameters (such as such as read-accept-infix-dot) are set to customize reading programs from strings and ports.

  This option is provided mainly for older test systems. Using make-module-evaluator with input starting with #lang is generally better.

• Finally, language can be a list whose first element is 'begin.

  In this case, a new namespace is created using sandbox-namespace-specs, which by default creates a new namespace using sandbox-make-namespace (which, in turn, uses make-base-namespace or make-gui-namespace depending on sandbox-gui-available and gui-available?).

  In the new namespace, language is evaluated as an expression to further initialize the namespace.

The requires list adds additional imports to the module or namespace for the input-programs, even in the case that require is not made available through the language.

The following examples illustrate the difference between an evaluator that puts the program in a module and one that merely initializes a top-level namespace:

> (define base-module-eval     ; a module cannot have free variables...     (make-evaluator 'racket/base '(define (f) later)))

program:1:0: later: unbound identifier in module

in: later

> (define base-module-eval     (make-evaluator 'racket/base '(define (f) later)                                  '(define later 5)))

> (base-module-eval '(f))

5

> (define base-top-eval     ; non-module code can have free variables:     (make-evaluator '(begin) '(define (f) later)))

> (base-top-eval '(+ 1 2))

3

> (base-top-eval '(define later 5))

> (base-top-eval '(f))

5

The make-module-evaluator function is essentially a restriction of make-evaluator, where the program must be a module, and all imports are part of the program. In some cases it is useful to restrict the program to be a module using a specific module in its language position — use the optional lang argument to specify such a restriction (the default, #f, means no restriction is enforced). When the program is specified as a path, then the path is implicitly added to the allow-for-load list.

(define base-module-eval2

; equivalent to base-module-eval:

(make-module-evaluator '(module m racket/base

(define (f) later)

(define later 5))))

The make-module-evaluator function can be convenient for testing module files: pass in a path value for the file name, and you get back an evaluator in the module’s context which you can use with your favorite test facility.

The allow-for-require and allow-for-load arguments adjust filesystem permissions to extend the set of files that are usable by the evaluator. Modules that are in a collection are automatically accessible, but the allow-for-require argument lists additional modules that can be required along with their imports (transitively) through a filesystem path. The allow-for-load argument similarly lists files that can be loaded. (The precise permissions needed for require versus load can differ.) The allow-read argument is for backward compatibility, only; each module-path? element of allow-read is effectively moved to allow-for-require, while other elements are moved to allow-for-load.

If the value of sandbox-propagate-exceptions is true (the default) when the sandbox is created, then exceptions (both syntax and run-time) are propagated as usual to the caller of the evaluation function (i.e., catch them with with-handlers). If the value of sandbox-propagate-exceptions is #f when the sandbox is created, then uncaught exceptions in a sandbox evaluation cause the error to be printed to the sandbox’s error port, and the caller of the evaluation receives #<void>.

Finally, the fact that a sandboxed evaluator accept syntax objects makes it usable as the value for current-eval, which means that you can easily start a sandboxed read-eval-print-loop. For example, here is a quick implementation of a networked REPL:

(define e (make-evaluator 'racket/base))

(let-values ([(i o) (tcp-accept (tcp-listen 9999))])

(parameterize ([current-input-port  i]

[current-output-port o]

[current-error-port  o]

[current-eval e])

(read-eval-print-loop)

(fprintf o "\nBye...\n")

(close-output-port o)))

Note that in this code it is only the REPL interactions that are going over the network connection; using I/O operations inside the REPL will still use the usual sandbox parameters (defaulting to no I/O). In addition, the code works only from an existing toplevel REPL — specifically, read-eval-print-loop reads a syntax value and gives it the lexical context of the current namespace. Here is a variation that uses the networked ports for user I/O, and works when used from a module (by using a new namespace):

(let-values ([(i o) (tcp-accept (tcp-listen 9999))])

(parameterize ([current-input-port   i]

[current-output-port  o]

[current-error-port   o]

[sandbox-input        i]

[sandbox-output       o]

[sandbox-error-output o]

[current-namespace (make-empty-namespace)])

(parameterize ([current-eval

(make-evaluator 'racket/base)])

(read-eval-print-loop))

(fprintf o "\nBye...\n")

(close-output-port o)))

14.12.1 Customizing Evaluators

The sandboxed evaluators that make-evaluator creates can be customized via many parameters. Most of the configuration parameters affect newly created evaluators; changing them has no effect on already-running evaluators.

The default configuration options are set for a very restricted sandboxed environment — one that is safe to make publicly available. Further customizations might be needed in case more privileges are needed, or if you want tighter restrictions. Another useful approach for customizing an evaluator is to begin with a relatively unrestricted configuration and add the desired restrictions. This approach is made possible by the call-with-trusted-sandbox-configuration function.

The sandbox environment uses two notions of restricting the time that evaluations takes: shallow time and deep time. Shallow time refers to the immediate execution of an expression. For example, a shallow time limit of five seconds would restrict (sleep 6) and other computations that take longer than five seconds. Deep time refers to the total execution of the expression and all threads and sub-processes that the expression creates. For example, a deep time limit of five seconds would restrict (thread (λ () (sleep 6))), which shallow time would not, as well as all expressions that shallow time would restrict. By default, most sandboxes only restrict shallow time to facilitate expressions that use threads.

Note that the reader function is usually called as is, but when it is used to read the program input for make-module-evaluator, read-accept-lang and read-accept-reader are set to #t.

• a string or byte string, which is converted to a port using open-input-string or open-input-bytes;

• an input port;

• the symbol 'pipe, which triggers the creation of a pipe, where put-input can return the output end of the pipe or write directly to it;

• a thunk, which is called to obtain a port (e.g., using current-input-port means that the evaluator input is the same as the calling context’s input).

• an output port, which is used as-is;

• the symbol 'bytes, which causes get-output to return the complete output as a byte string as long as the evaluator has not yet terminated (so that the size of the bytes can be charged to the evaluator);

• the symbol 'string, which is similar to 'bytes, but makes get-output produce a string;

• the symbol 'pipe, which triggers the creation of a pipe, where get-output returns the input end of the pipe;

• a thunk, which is called to obtain a port (e.g., using current-output-port means that the evaluator output is not diverted).

The default is (lambda () (dup-output-port (current-error-port))), which means that the error output of the generated evaluator goes to the calling context’s error port.

The default value is #f.

The default is (list sandbox-make-namespace).

The module paths are needed for sharing module instantiations between the sandbox and the caller. For example, sandbox code that returns posn values (from the lang/posn module) will not be recognized as such by your own code by default, since the sandbox will have its own instance of lang/posn and thus its own struct type for posns. To be able to use such values, include 'lang/posn in the list of module paths.

When testing code that uses a teaching language, the following piece of code can be helpful:

(sandbox-namespace-specs

(let ([specs (sandbox-namespace-specs)])

`(,(car specs)

,@(cdr specs)

lang/posn

,@(if (gui-available?) '(mrlib/cache-image-snip) '()))))

Various aspects of the library change when the GUI library is available, such as using a new eventspace for each evaluator.

The access mode symbol is one of: 'execute, 'write, 'delete, 'read, or 'exists. These symbols are in decreasing order: each implies access for the following modes too (e.g., 'read allows reading or checking for existence).

The path regexp is used to identify paths that are granted access. It can also be given as a path (or a string or a byte string), which is (made into a complete path, cleansed, simplified, and then) converted to a regexp that allows the path and sub-directories; e.g., "/foo/bar" applies to "/foo/bar/baz".

An additional mode symbol, 'read-bytecode, is not part of the linear order of these modes. Specifying this mode is similar to specifying 'read, but it is not implied by any other mode. (For example, even if you specify 'write for a certain path, you need to also specify 'read-bytecode to grant this permission.) The sandbox usually works in the context of a lower code inspector (see sandbox-make-code-inspector) which prevents loading of untrusted bytecode files — the sandbox is set-up to allow loading bytecode from files that are specified with 'read-bytecode. This specification is given by default to the Racket collection hierarchy (including user-specific libraries) and to libraries that are explicitly specified in an #:allow-read argument. (Note that this applies for loading bytecode files only, under a lower code inspector it is still impossible to use protected module bindings (see Code Inspectors).)

The default value is null, but when an evaluator is created, it is augmented by 'read-bytecode permissions that make it possible to use collection libraries (including sandbox-override-collection-paths). See make-evaluator for more information.

Note that (when memory accounting is enabled) memory is attributed to the highest custodian that refers to it. This means that if you inspect a value that sandboxed evaluation returns outside of the sandbox, your own custodian will be charged for it. To ensure that it is charged back to the sandbox, you should remove references to such values when the code is done inspecting it.

When limits are set, call-with-limits (see below) is wrapped around each use of the evaluator, so consuming too much time or memory results in an exception. Change the limits of a running evaluator using set-eval-limits.

A custodian’s limit is checked only after a garbage collection, except that it may also be checked during certain large allocations that are individually larger than the custodian’s limit.

The current-load/use-compiled handler is setup to allow loading of bytecode files under the original code inspector when sandbox-path-permissions allows it through a 'read-bytecode mode symbol, which makes loading libraries possible.

If the value is 'propagate (the default), then a new plumber is created, and a flush callback is added to the current plumber to propagate the request to the new plumber within the created sandbox (if the sandbox has not already terminated).

Added in version 6.0.1.8 of package sandbox-lib.

14.12.2 Interacting with Evaluators

The following functions are used to interact with a sandboxed evaluator in addition to using it to evaluate code.

Killing an evaluator is similar to sending an eof value to the evaluator, except that an eof value will raise an error immediately.

(One use for this custodian is with current-memory-use, where the per-interaction sub-custodians will not be charged with the memory for the whole sandbox.)

This procedure should be used to modify an existing evaluator limits, because changing the sandbox-eval-limits parameter does not affect existing evaluators. See also call-with-limits.

This procedure should be used to modify an existing evaluator handler, because changing the sandbox-eval-handlers parameter does not affect existing evaluators. See also call-with-custodian-shutdown and call-with-killing-threads for two useful handlers that are provided.

• if it was 'pipe, then get-output returns the input port end of the created pipe;

• if it was 'bytes or 'string, then the result is the accumulated output, and the output port is reset so each call returns a different piece of the evaluator’s output (note that results are available only until the evaluator has terminated, and any allocations of the output are subject to the sandbox memory limit);

• otherwise, it returns #f.

The prog? argument specifies whether to obtain expressions that were uncovered after only the original input program was evaluated (#t) or after all later uses of the evaluator (#f). Using #t retrieves a list that is saved after the input program is evaluated, and before the evaluator is used, so the result is always the same.

A #t value of prog? is useful for testing student programs to find out whether a submission has sufficient test coverage built in. A #f value is useful for writing test suites for a program to ensure that your tests cover the whole code.

The second optional argument, src, specifies that the result should be filtered to hold only syntax objects whose source matches src. The default is the source that was used in the program code, if there was one. Note that 'program is used as the source value if the input program was given as S-expressions or as a string (and in these cases it will be the default for filtering). If given #f, the result is the unfiltered list of expressions.

The resulting list of syntax objects has at most one expression for each position and span. Thus, the contents may be unreliable, but the position information is reliable (i.e., it always indicates source code that would be painted red in DrRacket when coverage information is used).

Note that if the input program is a sequence of syntax values, either make sure that they have 'program as the source field, or use the src argument. Using a sequence of S-expressions (not syntax objects) for an input program leads to unreliable coverage results, since each expression may be assigned a single source location.

This process is usually similar to (evaluator (list thunk)), except that it does not rely on the common meaning of a sexpr-based syntax with list expressions as function application (which is not true in all languages). Note that this is more useful for meta-level operations such as namespace manipulation, it is not intended to be used as a safe-evaluation replacement (i.e., using the sandbox evaluator as usual).

14.12.3 Miscellaneous

The value of gui? is no longer used by racket/sandbox itself. Instead, gui-available? and sandbox-gui-available are checked at the time that a sandbox evaluator is created.

Sandboxed evaluators use call-with-limits, according to the sandbox-eval-limits setting and uses of set-eval-limits: each expression evaluation is protected from timeouts and memory problems. Use call-with-limits directly only to limit a whole testing session, instead of each expression.