Re: [RFC 2/2] signal: extend pidfd_send_signal() to allow expedited process killing

[Daniel Colascione]

On Thu, Apr 11, 2019 at 10:36 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Apr 11, 2019 at 10:33:32AM -0700, Daniel Colascione wrote:
> > On Thu, Apr 11, 2019 at 10:09 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > > On Thu, Apr 11, 2019 at 8:33 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Apr 10, 2019 at 06:43:53PM -0700, Suren Baghdasaryan wrote:
> > > > > Add new SS_EXPEDITE flag to be used when sending SIGKILL via
> > > > > pidfd_send_signal() syscall to allow expedited memory reclaim of the
> > > > > victim process. The usage of this flag is currently limited to SIGKILL
> > > > > signal and only to privileged users.
> > > >
> > > > What is the downside of doing expedited memory reclaim?  ie why not do it
> > > > every time a process is going to die?
> > >
> > > I think with an implementation that does not use/abuse oom-reaper
> > > thread this could be done for any kill. As I mentioned oom-reaper is a
> > > limited resource which has access to memory reserves and should not be
> > > abused in the way I do in this reference implementation.
> > > While there might be downsides that I don't know of, I'm not sure it's
> > > required to hurry every kill's memory reclaim. I think there are cases
> > > when resource deallocation is critical, for example when we kill to
> > > relieve resource shortage and there are kills when reclaim speed is
> > > not essential. It would be great if we can identify urgent cases
> > > without userspace hints, so I'm open to suggestions that do not
> > > involve additional flags.
> >
> > I was imagining a PI-ish approach where we'd reap in case an RT
> > process was waiting on the death of some other process. I'd still
> > prefer the API I proposed in the other message because it gets the
> > kernel out of the business of deciding what the right signal is. I'm a
> > huge believer in "mechanism, not policy".
>
> It's not a question of the kernel deciding what the right signal is.
> The kernel knows whether a signal is fatal to a particular process or not.
> The question is whether the killing process should do the work of reaping
> the dying process's resources sometimes, always or never.  Currently,
> that is never (the process reaps its own resources); Suren is suggesting
> sometimes, and I'm asking "Why not always?"

FWIW, Suren's initial proposal is that the oom_reaper kthread do the
reaping, not the process sending the kill. Are you suggesting that
sending SIGKILL should spend a while in signal delivery reaping pages
before returning? I thought about just doing it this way, but I didn't
like the idea: it'd slow down mass-killing programs like killall(1).
Programs expect sending SIGKILL to be a fast operation that returns
immediately.