[mcc0nnell]

Summary

Fixes #369.

doBindMountAt() previously stripped MS_RDONLY before applyMountFlags(), so recursive mount_setattr(AT_RECURSIVE) propagated nosuid, nodev, and noexec but not read-only. The later remountPt() affects only the top-level bind mount. When is_root_rw: true skips the root-wide readonly pass, a nested submount could remain writable.

This change:

• Applies MS_RDONLY to the cloned bind subtree before move_mount(); AT_RECURSIVE is selected when MS_REC is present.
• Fails closed if applying read-only attributes fails, rather than degrading to a top-level-only remount.
• Adds a privilege-gated regression with a real nested tmpfs and a make target. It skips when mount privileges are unavailable.

Validation

Reproduced on Linux 6.8 in an isolated privileged Docker mount namespace, explicitly forcing --experimental_mnt=new and writable jail root (--rw):

Test	Upstream master	Patch
Readonly recursive bind, top-level	ro	ro
Readonly recursive bind, nested tmpfs	rw (bug)	ro
Writable recursive bind, top-level and nested tmpfs	—	Both rw, writes allowed
Readonly file bind	—	ro, write rejected

The exact committed regression (make test-newapi-recursive-readonly) fails against upstream master (exit 1) and passes against the patched commit (exit 0). With insufficient privileges, the test reports a skip.

Additional checks against the exact commit:

• make -j4 nsjail — exit 0.
• make test-cmdline — 7/7 pass.
• Four native test binaries — all exit 0.
• git diff --check — clean.

The full network-dependent make test suite was not executed. Real mount behavior was verified as described above.

Only mnt_newapi.cc, tests/newapi_recursive_readonly.sh, and Makefile changed.

[robertswiecki]

Thank you!