Linux 5.15 Adds New Syscall To More Quickly Free Memory Of Dying Processes

[phoronix]

Phoronix: Linux 5.15 Adds New Syscall To More Quickly Free Memory Of Dying Processes

To help out memory pressure / out-of-memory killing solutions like systemd-oomd or Android's LMKD, Linux 5.15 is introducing the "process_mrelease" system call to more quickly free the memory of dying processes...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.15-process-mrelease

[pipe13]

Not even properly dead and already harvesting its organs

[sinepgib]

Not even properly dead and already harvesting its organs

Well, you know, the fresher the better.

[boxie]

What happens to the dying process? does it get to run any code? is execution immediately halted?

[pal666]

What happens to the dying process? does it get to run any code? is execution immediately halted?

subj talks about process sleeping in uninterruptible sleep, it doesn't run anything, just waits

[StarterX4]

And what about Zombie processes? Another syscall would be needed.

[hakavlad]

And what about Zombie processes? Another syscall would be needed.

VmRSS=0 if state=Z, no syscall needed.

[hakavlad]

What happens to the dying process? does it get to run any code? is execution immediately halted?

does it get to run any code? - no

is execution immediately halted? - yes

Memory is not released instantly, but `/proc/pid/statm` shows 0 0 0 0 0 immediately after sending SIGKILL to the process.

Memory can be released for a few seconds if the process is large and some of its memory is in the swap space.

[boxie]

does it get to run any code? - no

is execution immediately halted? - yes

Memory is not released instantly, but `/proc/pid/statm` shows 0 0 0 0 0 immediately after sending SIGKILL to the process.

Memory can be released for a few seconds if the process is large and some of its memory is in the swap space.

thank you for a very nice explanation!

[carewolf]

Just fix the restore from starting to use swap memory, it is what is killing everything that Linux never recovers from a swap condition. It doesn't try, and if you force it by disabling swap, it takes 10-100x longer than it should.