Skip to content

A lightweight process isolation tool, requiring absolutely no privileges to run

License

Notifications You must be signed in to change notification settings

rapidlua/sandals

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

71 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

sandals

A lightweight process isolation tool for Linux. It is built using Linux namespaces, cgroups v2, and seccomp-bpf syscall filters.

$ echo '{"cmd":["ps","-A"],
         "pipes":[{"dest":"/dev/stdout","stdout":true}],
         "mounts":[{"type":"proc","dest":"/proc"}]
        }' | sandals
  PID TTY          TIME CMD
    1 ?        00:00:00 sandals
    2 ?        00:00:00 ps
{"status":"exited","code":0}

Highlights

  1. speaks JSON — A departure from long-standing UNIX tradition, sandals reads JSON request from STDIN and writes JSON response to STDOUT, instead of relying on program arguments and exit code.
  2. detailed status — Response tells whether the task exited normally or was killed or terminated due to compromised integrity (see below).
  3. task integrity — Sandbox introduces new modes of failure, e.g. disk full error due to to filesystem quota in effect; e.g. a subprocess terminated due to exceeded memory limit. A task integrity is compromised; it is unlikely to recover and it might produce bogus results if not prepared to handle an unusual failure. The compromised task is terminated and detailed status tells the reason and gives a hint if certain limits should be increased.
  4. rootless — No privileges required. No suid binaries involved.
  5. as fast as possible — Wrapping a task in docker run reportedly adds 300ms overhead. Sandals reduces that to a mere 5ms, a boon for short-lived tasks!
  6. exposes Linux kernel features instead of inventing higher-level abstractions — Primarily a side effect of being fast and lean. We aren't Docker, the user should be sufficiently versed in namespaces, cgroups, seccomp and mounts.

Other mature lightweight sandboxes are readily available: isolate (est. 2013), firejail (est. 2014), nsjail (est. 2015) to name a few. They are similarly low-level and comparably fast (though we managed to innovatively squeeze almost 20ms by doing namespaced process shutdown asynchronously). Our other features are unparalleled though.

Most notably, existing solutions require elevated privileges in order to set up a constrained sandbox. This is an inherent limitation of cgroups v1. Sandals, on the other hand, takes advantage of cgroups v2.

Installation

$ make && make install

User guide

(This guide is not meant to be exhaustive, check Reference for further details.)

Sandals runs untrusted code in a private set of Linux namespaces. Namespaces partition system objects. E.g. IPC namespaces partition IPC objects. Sandals creates the full set of private namespaces by default. Sandboxed code is unable to interact with host processes, either directly via signals or indirectly via IPC or sockets. It has no access to the network either.

Restricting access to the filesystem

TODO

Limiting resource usage with cgroups

TODO

Reference

Response

JSON object with status key. Depending on the status aditional keys might be present:

  • exited: exited normally
    • code: process exit code
  • killed: killed by signal
    • signal: killing signal (string, e.g. SIGTERM)
  • memoryLimit: memory limit as set by cgroup's memory.max exceeded
  • pidsLimit: pids limit as set by cgroup's pids.max exceeded
  • timeLimit: run time limit exceeded, see timeLimit in Request section
  • outputLimit: task output size limit exceeded, see pipes, copyFiles and stdStreams in Request section
  • requestInvalid: request JSON invalid
    • description: error description
  • internalError: internal error
    • description: error description

Request

JSON object with the mandatory cmd key. Ex: {"cmd":["uname","-a"]}.

Optional keys:

  • hostName: string

    Host name as observed inside sandbox. Default: sandals.

  • domainName: string

    Domain name as observed inside sandbox. Default: sandals.

  • uid: number

    User id as observed inside sandbox. Default: 0.

  • gid: number

    Group id as observed inside sandbox. Default: 0.

  • chroot: string

    A path to use as filesystem root inside sandbox. Default: /.

  • mounts: object []

    A list of mounts to apply to filesystem view inside sandbox. Default: [].

    Every mount has the mandatory type key. Type is either bind for a bind mount or a value recognized by mount system call, like tmpfs or proc.

    The mandatory dest key tells the destination of the mount. Chroot prefix (if any) is automatically prepended to the given path.

    Bind mounts must specify the source path with src key.

    Optional options string (empty by default) is passed to mount syscall verbatim.

    Optional ro boolean key (default: false) turns on read-only mode for the mount.

  • cgroupConfig: object

    Enable resource limiting with cgroups. Default: cgroups not used.

    Ex: {"memory.max": "1000000", "memory.swap.max": "1000000"}. For each key/value pair, the value is written to the file named by key in the cgroup directory.

  • cgroup, cgroupRoot: string

    If cgroupConfig is present, sandboxed task is put into a separate cgroup. This is either an existing cgroup as specified by cgroup key or a new one if the later key is absent.

    A new cgroup is created under cgroupRoot if present, otherwize a new cgroup is spawned as a sibling of the current cgroup.

    If a new cgroup was created it is removed when sandals exits.

  • seccompPolicy: string

    A syscall filtering policy in Kafel syntax. Filtering disabled by default.

  • vaRandomize: boolean

    ASLR, enabled by default.

  • env: string []

    Task's environment as a list of KEY=VALUE strings. Empty by default.

  • workDir: string

    Task's working directory. Default: /.

    Chroot prefix (if any) is automatically prepended to the given path.

  • timeLimit: number

    A time limit in seconds. No limit by default.

  • pipes: object []

    A list of unidirectional channels for streaming data out of the sandbox.

    Mandatory dest key names the destination file to write the data.

    At least one of stdout, stderr or src keys must be present. If stdout key is set to true the pipe is attached as a task's standard output. Set stderr to true to attach the standard error stream. The last option is to expose the pipe ingress as a fifo in the filesystem. Chroot prefix (if any) is automatically prepended to the path in src key.

    Optional limit numeric key caps the maximum amount of collected data (no limit by default). If the limit is exeeded, the task terminates with status:outputLimit.

  • copyFiles: object []

    A list of files to copy out of the sandbox once it terminates.

    Subkeys are the same as in pipe object, see pipes.

  • stdStreams: object

    Subkeys (same meaning as in pipe object, see pipes):

    • dest

    • limit

    Capture both stdout and stderr simultaneously. Use case: present task's output exactly as it would appear if invoked in a terminal. Style stderr content differently (eg: red color).

    Every chunk of data is prefixed with 32-bit integer in the network byte order. Bits 0..30 encode the chunk length. Bit 31 tells whether the chunk origin is stdout (0) or stderr (1).

    Note: if size limit is exceeded the last packet (header + chunk) might be cut short.

    Note: capturing task's output exactly as it would appear if invoked in a terminal, while distinguishing stdout/stderr content, is tricky. In most programming languages, if a stream is writing to a terminal, a different buffering strategy is used. If the same pipe is attached to stdout and stderr, it won't be possible to tell which stream the particular octet originated from. If two pipes are used, preserving relative ordering of chunks is challenging.

    The current implementation utilizes UNIX datagram sockets, which solve the ordering problem nicely. The caveat is that any write to stdout/sdterr exceeding the maximum datagram size (~200KiB) will fail. Also streams buffering doesn't match.

    If the perfect solution is desired, we suggest installing a custom Linux kernel module (available separately). Sandals will use the kernel module if present.