Re: [PATCH 1/1] mm: introduce process_reap system call

[David Hildenbrand]

On 02.07.21 17:27, Christian Brauner wrote:
> On Thu, Jul 01, 2021 at 03:59:48PM -0700, Suren Baghdasaryan wrote:
> > On Wed, Jun 30, 2021 at 5:44 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> > > On Wed, Jun 30, 2021 at 2:45 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> > > > On Wed, Jun 30, 2021 at 11:51:36AM -0700, Suren Baghdasaryan wrote:
> > > > > On Wed, Jun 30, 2021 at 11:26 AM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> > > > > > Also, please consider removing all mention of the word "reap" from the
> > > > > > user API.  For better or for worse, "reap" in UNIX refers to what
> > > > > > happens when a dead task gets wait()ed.  I sincerely wish I could go
> > > > > > back in time and gently encourage whomever invented that particular
> > > > > > abomination to change their mind, but my time machine doesn't work.
> > > > >
> > > > > I see. Thanks for the note. How about process_mem_release() and
> > > > > replacing reap with release everywhere?
> > > >
> > > > I don't quite understand the objection. This syscall works on tasks
> > > > that are at the end of their life, right? Isn't something like
> > > > process_mreap() establishing exactly the mental link we want here?
> > > > Release is less descriptive for what this thing is to be used for.
> > >
> > > For better or for worse, "reap" means to make a zombie pid go away.
> > >  From the description, this new operation takes a dying process (not
> > > necessarily a zombie yet) and aggressively frees its memory.  This is
> > > a different optioneration.
> > >
> > > How about "free_dying_process_memory"?
> >
> > process_mreap sounds definitely better and in line with names like
> > process_madvise. So maybe we can use it?
>
> That one was my favorite from the list I gave too but maybe we can
> satisfy Andy too if we use one of:
> - process_mfree()
> - process_mrelease()

FWIW, I tend to like process_mrelease(), due to the implied "release" 
("free the memory if there are no other references") semantics. Further, 
a new syscall feels cleaner than some magic sysfs/procfs toggle. Just my 
2 cents.

--
Thanks,

David / dhildenb