pub fn try_lock_memory()Expand description
Pin engine pages into RAM with mlockall, lock-on-fault.
Goal: prevent the kernel from faulting out engine pages under memory pressure, which otherwise surfaces as 100µs–10ms tail spikes.
We pass MCL_ONFAULT alongside MCL_CURRENT | MCL_FUTURE deliberately.
This runs early — before the application’s state and the rest of the
pipeline are built — so MCL_FUTURE is required to cover those later
allocations.
But plain MCL_FUTURE marks every future mapping VM_LOCKED and makes
the kernel eagerly populate the whole mapping at mmap time, inside an
uninterruptible __mm_populate walk. Any sizeable post-init allocation
(a large malloc glibc serves via mmap, a snapshot/clone) then stalls
the allocating thread until every page is faulted in — long enough to
even outlast SIGKILL. MCL_ONFAULT switches future mappings to
lock-on-fault: pages are locked as they are touched, not populated up
front. The fault-out protection we actually want is preserved (a page,
once touched on the hot path, stays resident), without the populate stall.
Best-effort: requires CAP_IPC_LOCK (or root) and a sufficient
RLIMIT_MEMLOCK. Logs a warning and continues on failure.