sys::open still sent the FULL path in a single SYS_OPENAT_INTO on the
namespace fd — the pre-fd-allocation contract that forced the namespace
manager to proxy every open with a blocking openat into the provider
(the mini boot-freeze head-of-line; see base initnsmgr commit adopting
the upstream capability-handoff path).
Port the upstream two-step (verbatim from upstream/master, matching the
open_into_upper and unlink helpers below that were already synced):
1. openat(ns_fd, full_path, O_DIRECTORY|O_CLOEXEC) — the nsmgr resolves
only the scheme part and dups the scheme-root capability.
2. openat(root_fd, reference, flags) — the kernel routes straight to the
provider; the nsmgr is out of the data path.
With this, a busy/slow provider can no longer stall unrelated opens
through the nsmgr's single request loop.
vasilito
2fa1aac513