process_mrelease(2)

Purpose

process_mrelease() asks the kernel to release the memory of a process that has already been selected to die, without waiting for the OOM reaper or the exiting thread to get there on its own. The caller holds a pidfd (pidfd_open) on the target and the kernel reaps the target’s address space the way the OOM reaper would (mm/oom_kill.c, SYSCALL_DEFINE2(process_mrelease)). It exists so that userspace OOM killers (systemd-oomd, oomd, container runtimes) can turn “kill” into “kill and reclaim immediately” instead of racing the victim’s exit path.

Synopsis

/* no glibc declaration; called directly */
long syscall(SYS_process_mrelease, int pidfd, unsigned int flags);

Arguments and behavior

pidfd is a file descriptor referring to the target process (pidfd conventions; PIDFD_SIGNAL applies). flags must be 0; any other value fails with EINVAL (mm/oom_kill.c, kernel constant check if (flags) return -EINVAL;). The kernel looks the task up from the pidfd, locks a thread that still has an mm_struct reference during group exit, and decides: if task_will_free_mem(p) is true the memory is reaped immediately; otherwise the call fails with EINVAL unless the work was already done (MMF_OOM_SKIP set on the mm) (mm/oom_kill.c, body). Reaping runs under mmap_read_lock_killable(), so a blocked caller can be interrupted with EINTR, and the MMF_OOM_SKIP check is repeated under the read lock to close the race with exit_mmap (commit 337546e83f, “prevent a race between process_mrelease and exit_mmap”).

Return value and errors

0 on success (memory of the dying process reclaimed or release work done); a negated error on the raw syscall path (no wrapper to translate it): EINVAL (flags nonzero, or the task is not going to die and the reaper has not run), ESRCH (the task no longer has an mm – already exited), EINTR (mmap read lock acquisition killed), plus pidfd lookup errors from pidfd_get_task (EBADF for a bad descriptor) (mm/oom_kill.c, body).

Flags and constants

None (flags must be 0).

Security and notes

Permission model: the caller needs to signal the target the way pidfd_send_signal does (the pidfd was obtained with the appropriate access; kernel checks ptrace_may_access-style access on the pidfd lookup). The design goal from the merge commit 884a7e5964: userspace OOM killing should not leave the victim’s memory pinned for long; on Android this cut low-memory-killer reclaim latency substantially. The companion semantic detail worth remembering: process_mrelease() never kills – the target must already be dying (SIGKILL delivered via pidfd_send_signal or the OOM killer); otherwise EINVAL. This page is source-limited: the behavior above is read from mm/oom_kill.c (master, fall 2026) because no man page was reachable; a future man2/process_mrelease.2 should be reconciled against it. Cross-reference: pidfd_open, pidfd_send_signal, exit_group, process_madvise.

References

• Kernel source: mm/oom_kill.c (SYSCALL_DEFINE2(process_mrelease), torvalds/linux master, fetched 2026-10-17)
• Merge commit: 884a7e5964 “mm: introduce process_mrelease system call” (2021-09-02, 5.15); race fix 337546e83f (2021-10-28)
• https://man7.org/linux/man-pages/man2/pidfd_open.2.html (pidfd conventions)
• worklog: no glibc wrapper symbol (nm -D, glibc 2.41)