Re: [PATCH v2] epoll: add nsec timeout support

[Andrew Morton]

On Mon, 16 Nov 2020 18:51:16 -0500 Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx> wrote:

> On Mon, Nov 16, 2020 at 6:36 PM Willem de Bruijn
> <willemdebruijn.kernel@xxxxxxxxx> wrote:
> >
> > On Mon, Nov 16, 2020 at 3:04 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, 16 Nov 2020 11:10:01 -0500 Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > >
> > > > From: Willem de Bruijn <willemb@xxxxxxxxxx>
> > > >
> > > > Add epoll_create1 flag EPOLL_NSTIMEO. When passed, this changes the
> > > > interpretation of argument timeout in epoll_wait from msec to nsec.
> > > >
> > > > Use cases such as datacenter networking operate on timescales well
> > > > below milliseconds. Shorter timeouts bounds their tail latency.
> > > > The underlying hrtimer is already programmed with nsec resolution.
> > >
> > > hm, maybe.  It's not very nice to be using one syscall to alter the
> > > interpretation of another syscall's argument in this fashion.  For
> > > example, one wonders how strace(1) is to properly interpret & display
> > > this argument?
> > >
> > > Did you consider adding epoll_wait2()/epoll_pwait2() syscalls which
> > > take a nsec timeout?  Seems simpler.
> >
> > I took a first stab. The patch does become quite a bit more complex.
> 
> Not complex in terms of timeout logic. Just a bigger patch, taking as
> example the recent commit ecb8ac8b1f14 that added process_madvise.

That's OK - it's mainly syscall table patchery.  The fs/ changes are
what matters.  And the interface.

> > I was not aware of how uncommon syscall argument interpretation
> > contingent on internal object state really is. Yes, that can
> > complicate inspection with strace, seccomp, ... This particular case
> > seems benign to me. But perhaps it sets a precedent.
> >
> > A new nsec resolution epoll syscall would be analogous to pselect and
> > ppoll, both of which switched to nsec resolution timespec.
> >
> > Since creating new syscalls is rare, add a flags argument at the same time?

Adding a syscall is pretty cheap - it's just a table entry.

> >
> > Then I would split the change in two: (1) add the new syscall with
> > extra flags argument, (2) define flag EPOLL_WAIT_NSTIMEO to explicitly
> > change the time scale of the timeout argument. To avoid easy mistakes
> > by callers in absence of stronger typing.

I don't understand this.  You're proposing that the new epoll_pwait2() be
able to take either msec or nsec, based on the flags argument?  With a
longer-term plan to deprecate the old epoll_pwait()?

If so, that's not likely to be viable - how can we ever know that the
whole world stopped using the old syscall?

> Come to think of it, better to convert to timespec to both have actual
> typing and consistency with ppoll/pselect.

Sure.

> > epoll_wait is missing from include/uapi/asm-generic/unistd.h as it is
> > superseded by epoll_pwait. Following the same rationale, add
> > epoll_pwait2 (only).

Sure.

> A separate RFC patch against manpages/master sent at the same time?

That's the common approach - a followup saying "here's what I'll send
to the manpages people if this gets merged".

And something under tools/testing/sefltests/ would be nice, if only so
that the various arch maintainers can verify that their new syscall is
working correctly.  Perhaps by adding a please-use-epoll_pwait2 arg to
the existing
tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c, if that
looks like a suitable testcase.