[RFC] Proposal: Kernel-supported mechanism for shared memory state cleanup on exit

[RFC] Proposal: Kernel-supported mechanism for shared memory state cleanup on exit

[Yongfeng Wang]

Hi Andrew Morton,

I’m looking for your feedback on a gap in the current synchronization
primitives: the lack of a general, kernel-supported mechanism to clean
up user-space shared memory state (e.g., refcounts, ownership flags)
when a process exits unexpectedly (SIGKILL/crash).

The Problem: While robust_list handles mutexes and SEM_UNDO handles
System V semaphores, there is no generic way to perform simple,
reversible operations (like clearing a bit or decrementing a counter)
on arbitrary shared memory upon process death. User-space recovery is
often unreliable in crash scenarios.

The Idea: A syscall or prctl extension allowing a process to register
a limited number of "deferred actions" on shared memory addresses.

Target: Small, management-plane state.

Operations: Atomic write/add/clear.

Execution: Handled by the kernel during do_exit(), ensuring atomicity
even if the process is killed.

Questions:

Do you see this as a problem worth solving at the kernel level, or
should this remain strictly in user-space?

Are there fundamental architectural blockers (e.g., page fault
handling during do_exit) that make this impractical?

I have a more detailed technical draft ready if this is a direction
you’d consider.

Best regards,

Yongfeng Wang