Skip to content

Feature: write-only / append-only filesystem grants (fs_write_only / fs_append) #150

Description

@dzerik

Summary

Add write-only (and ideally append-only) filesystem grants to the public API, e.g. fs_write_only(path) / fs_append(path), in addition to the existing fs_read/fs_write`.

Today fs_write grants read and write: a path passed to fs_write is also readable (the write mask is built as READ_ACCESS | write/create/delete flags).
There is no way to let a sandboxed process write to a location it cannot read back, nor to enforce append-only (no truncate/overwrite of existing content).

Motivation

A supervisor/host frequently needs to hand a sandboxed workload a location for significant or audit-relevant output — logs, audit records, telemetry, a drop-box — that the workload must be able to append to but must not be
able to:

  • read back (confidentiality / no exfiltration of other tenants' or prior records that share the sink), or
  • overwrite / truncate existing records (tamper-evidence — the workload cannot rewrite history).

With the current read+write grant, sandboxed code that knows the path can open the sink for reading or O_TRUNC and destroy/alter earlier records. For a product exposed to untrusted external code, a tamper-evident audit trail is a
common requirement, and today it can only be approximated by writing the sink through a separate privileged channel rather than the filesystem.

Why this is a good fit for Landlock

Landlock already models these as distinct rights, and sandlock already computes per-grant masks (crates/sandlock-core/src/landlock.rs):

  • write-only: LANDLOCK_ACCESS_FS_MAKE_REG | LANDLOCK_ACCESS_FS_WRITE_FILE (+ directory traversal to reach the path), without READ_FILE / READ_DIR. open(path, O_RDONLY) from the sandbox then fails with EACCES.
  • append-only: additionally withhold LANDLOCK_ACCESS_FS_TRUNCATE so the sandbox cannot O_TRUNC/ftruncate existing content (combined with the host opening/handing the sink such that writes are appends).

So the enforcement is kernel-backed and needs no new supervisor logic — only a new grant constructor that emits the narrower mask instead of the current read+write one.

Proposed API (sketch)

// builder
.fs_write_only(path)   // MAKE_REG + WRITE_FILE (+ traversal), no READ_FILE/READ_DIR
.fs_append(path)       // as write_only, also no TRUNCATE

with the same surfaced through the Python/Go bindings (a write_only / append list alongside fs_readable / fs_writable).

Notes

  • Independent of Sandbox.run() deadlocks when the child writes more than the pipe buffer to stdout/stderr #149 (the run() stdout-capture deadlock); this is a separate confinement-granularity feature.
  • A sandboxed process that opened the file itself still holds whatever fd it opened, so write-only is meaningful primarily for sinks the process must open fresh (it cannot obtain a read fd) or that the supervisor pre-creates — the point is that the path grants no read/truncate right, so no read/rewind/ truncate is possible via the filesystem.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions