Re: [RFC] simple_lmk: Introduce Simple Low Memory Killer for Android

[Christian Brauner]

On Fri, Mar 15, 2019 at 02:13:24PM -0400, Joel Fernandes wrote:
> On Fri, Mar 15, 2019 at 07:03:07PM +0100, Christian Brauner wrote:
> > On Thu, Mar 14, 2019 at 09:36:43PM -0700, Daniel Colascione wrote:
> > > On Thu, Mar 14, 2019 at 8:16 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, 14 Mar 2019 13:49:11 -0700
> > > > Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > > Perhaps I'm missing something, but if you want to know when a process has died
> > > > > after sending a SIGKILL to it, then why not just make the SIGKILL optionally
> > > > > block until the process has died completely? It'd be rather trivial to just
> > > > > store a pointer to an onstack completion inside the victim process' task_struct,
> > > > > and then complete it in free_task().
> > > >
> > > > How would you implement such a method in userspace? kill() doesn't take
> > > > any parameters but the pid of the process you want to send a signal to,
> > > > and the signal to send. This would require a new system call, and be
> > > > quite a bit of work.
> > > 
> > > That's what the pidfd work is for. Please read the original threads
> > > about the motivation and design of that facility.
> > > 
> > > > If you can solve this with an ebpf program, I
> > > > strongly suggest you do that instead.
> > > 
> > > Regarding process death notification: I will absolutely not support
> > > putting aBPF and perf trace events on the critical path of core system
> > > memory management functionality. Tracing and monitoring facilities are
> > > great for learning about the system, but they were never intended to
> > > be load-bearing. The proposed eBPF process-monitoring approach is just
> > > a variant of the netlink proposal we discussed previously on the pidfd
> > > threads; it has all of its drawbacks. We really need a core system
> > > call  --- really, we've needed robust process management since the
> > > creation of unix --- and I'm glad that we're finally getting it.
> > > Adding new system calls is not expensive; going to great lengths to
> > > avoid adding one is like calling a helicopter to avoid crossing the
> > > street. I don't think we should present an abuse of the debugging and
> > > performance monitoring infrastructure as an alternative to a robust
> > > and desperately-needed bit of core functionality that's neither hard
> > > to add nor complex to implement nor expensive to use.
> > > 
> > > Regarding the proposal for a new kernel-side lmkd: when possible, the
> > > kernel should provide mechanism, not policy. Putting the low memory
> > > killer back into the kernel after we've spent significant effort
> > > making it possible for userspace to do that job. Compared to kernel
> > > code, more easily understood, more easily debuggable, more easily
> > > updated, and much safer. If we *can* move something out of the kernel,
> > > we should. This patch moves us in exactly the wrong direction. Yes, we
> > > need *something* that sits synchronously astride the page allocation
> > > path and does *something* to stop a busy beaver allocator that eats
> > > all the available memory before lmkd, even mlocked and realtime, can
> > > respond. The OOM killer is adequate for this very rare case.
> > > 
> > > With respect to kill timing: Tim is right about the need for two
> > > levels of policy: first, a high-level process prioritization and
> > > memory-demand balancing scheme (which is what OOM score adjustment
> > > code in ActivityManager amounts to); and second, a low-level
> > > process-killing methodology that maximizes sustainable memory reclaim
> > > and minimizes unwanted side effects while killing those processes that
> > > should be dead. Both of these policies belong in userspace --- because
> > > they *can* be in userspace --- and userspace needs only a few tools,
> > > most of which already exist, to do a perfectly adequate job.
> > > 
> > > We do want killed processes to die promptly. That's why I support
> > > boosting a process's priority somehow when lmkd is about to kill it.
> > > The precise way in which we do that --- involving not only actual
> > > priority, but scheduler knobs, cgroup assignment, core affinity, and
> > > so on --- is a complex topic best left to userspace. lmkd already has
> > > all the knobs it needs to implement whatever priority boosting policy
> > > it wants.
> > > 
> > > Hell, once we add a pidfd_wait --- which I plan to work on, assuming
> > > nobody beats me to it, after pidfd_send_signal lands --- you can
> > 
> > Daniel,
> > 
> > I've just been talking to Joel.
> > I actually "expected" you to work pidfd_wait() after prior
> > conversations we had on the pidfd_send_signal() patchsets. :) That's why
> > I got a separate git tree on kernel.org since I expect a lot more work
> > to come. I hope that Linus still decides to pull pidfd_send_signal()
> > before Sunday (For the ones who have missed the link in a prior response
> > of mine:
> > https://lkml.org/lkml/2019/3/12/439
> > 
> > This is the first merge window I sent this PR.
> > 
> > The pidfd tree has a branch for-next that is already tracked by Stephen
> > in linux-next since the 5.0 merge window. The patches for
> > pidfd_send_signal() sit in the pidfd branch.
> > I'd be happy to share the tree with you and Joel (We can rename it if
> > you prefer I don't care).
> > I would really like to centralize this work so that we sort of have a
> > "united front" and end up with a coherent api and can send PRs from a
> > centralized place:
> > https://git.kernel.org/pub/scm/linux/kernel/git/brauner/linux.git/
> 
> I am totally onboard with working together / reviewing this work with you all
> on a common tree somewhere (Christian's pidfd tree is fine). I was curious,

Excellent.

> why do we want to add a new syscall (pidfd_wait) though? Why not just use
> standard poll/epoll interface on the proc fd like Daniel was suggesting.
> AFAIK, once the proc file is opened, the struct pid is essentially pinned
> even though the proc number may be reused. Then the caller can just poll.
> We can add a waitqueue to struct pid, and wake up any waiters on process
> death (A quick look shows task_struct can be mapped to its struct pid) and
> also possibly optimize it using Steve's TIF flag idea. No new syscall is
> needed then, let me know if I missed something?

Huh, I thought that Daniel was against the poll/epoll solution?
I have no clear opinion on what is better at the moment since I have
been mostly concerned with getting pidfd_send_signal() into shape and
was reluctant to put more ideas/work into this if it gets shutdown.
Once we have pidfd_send_signal() the wait discussion makes sense.

Thanks!
Christian