[Alb3e3]

Fixes #125.

Summary

--bindmount_ro / -R uses MS_REC while creating the bind, but the final MS_REMOUNT | MS_BIND | MS_RDONLY operation only changes the top mount. A writable nested mount such as /dev/shm therefore remains writable inside an otherwise read-only recursive bind.

• Apply MOUNT_ATTR_RDONLY with mount_setattr(2) and AT_RECURSIVE before the existing top-level remount.
• Use the same helper for the legacy and new mount backends.
• Keep the existing ordered per-mount remount pass, so a later explicit writable child bind still overrides the recursive parent policy.
• Fail sandbox setup when final mount attributes cannot be enforced instead of continuing with a weaker mount policy.

On kernels without mount_setattr(2) support, recursive read-only bind setup now fails closed rather than silently exposing writable nested mounts.

Verification

• Confirmed the regression on the unmodified image with both --experimental_mnt=old and --experimental_mnt=new: -R /dev still allowed creating a file in /dev/shm.
• Confirmed both patched backends reject the nested write with Read-only file system.
• Confirmed a later -B /dev/shm remains writable with both backends.
• Built with make -j2 and the project Dockerfile.

[Alb3e3]

Rebased this branch onto current master and reran the focused recursive-bind checks on both mount backends. -R /dev now blocks nested /dev/shm writes with Read-only file system, while an explicit child override with -B /dev/shm stays writable in both --experimental_mnt=old and --experimental_mnt=new. Docker build also still passes. Would appreciate another look when you have a moment.

[robertswiecki]

This has been superseded by #283 / #284 which goes a bit further (more flags)

[Alb3e3]

Makes sense. #284 applying the per-mount nosuid/nodev/noexec recursively is the more complete fix. Thanks for the pointer, and good to see the recursive-bind hardening land.