Harden actiondfs and isolate remote action execution

[dzbarsky]

Summary

• Harden actiondfs directory parsing, inode identity, staged-directory streaming, symlink handling, metadata synchronization, backing-file validation, and inode/dentry lifetimes.
• Isolate each remote action with mount/network/PID namespaces, cgroups, seccomp filters, restricted BPF/vsock/io_uring entry points, execution deadlines, and process-tree cancellation.
• Bound protobuf decoding, HTTP/2 headers/bodies/concurrent streams, CAS and ByteStream transfers, ActionCache results, and declared/generated outputs.
• Extend end-to-end coverage for staged filesystem operations, cgroup PID limits, sandbox isolation, cancellation, timeout recovery, malformed protobufs, and unsafe CAS entries.
• Keep Linux LLVM benchmark instrumentation opt-in and avoid x86 seccomp BPF JIT compilation without disabling seccomp filters.

Validation

• Remote Bazel build of 66 actual repository targets, including both Linux kernels and release binaries; excludes 12 upstream-generated Linux kernel host tools that already fail as standalone macOS targets on main.
• Remote Bazel tests: //src:unit_tests and //tools:e2e_action_tool_tests.
• (cd test && bazel test //:parse_timings_test).
• Production tools/e2e.sh vm on macOS Virtualization.framework, including the full remote-execution stress workload, PID limits, execution deadlines, and recovery.
• Production x86 Linux QEMU LLVM smoke with 1,998 remote actions and 1,998 native Linux actions.

Linux CI performance comparison

Both runs use the same @llvm-project//llvm:llvm-tblgen workload, warmup, 4-vCPU/4-GiB x86 QEMU VM, default Bazel job count, production actiondfs configuration, remote cache behavior, and 1,998 executed actions.

• Pull request: actiond 280.135s, native Linux 282.035s, actiond/native 0.993×.
• Upstream main: actiond 309.659s, native Linux 314.245s, actiond/native 0.985×.

The normalized difference is 0.8 percentage points across separate GitHub-hosted runners; actiond is faster than native Linux in both measurements.

Windows CI performance comparison

Both Windows runs execute the same @llvm-project//llvm:llvm-tblgen workload with eight Bazel jobs, four VM CPUs, and 1,998 executed actions.

• Pull request: actiond 341.332s, native Windows 333.829s, actiond/native 1.022×.
• Upstream main: actiond 262.970s, native Windows 260.088s, actiond/native 1.011×.

The normalized difference is 1.1 percentage points across separate GitHub-hosted runners; an earlier pull-request run measured 0.975×, demonstrating the expected runner variation.