On Wed, 31 Jul 2019, Andy Lutomirski wrote:
> On Mon, Jul 29, 2019 at 2:58 PM Dmitry Safonov <dima@xxxxxxxxxx> wrote:
> > As it has been discussed on timens RFC, adding a new conditional branch
> > `if (inside_time_ns)` on VDSO for all processes is undesirable.
> >
> > Addressing those problems, there are two versions of VDSO's .so:
> > for host tasks (without any penalty) and for processes inside of time
> > namespace with clk_to_ns() that subtracts offsets from host's time.
> >
> > The timens code in vdso looks like this:
> >
> > if (timens_static_branch_unlikely()) {
> > clk_to_ns(clk, ts);
> > }
>
> I'm confused. Now we effectively have *three* versions: the vDSO
> without timens, and vDSO with timens but with it switched off, and the
> vDSO with timens on. This seems like too much.
The problem is that if you have a single VDSO, then omce one process joins
a time namespace _ALL_ other processes get an extra conditional with at
least one extra cache line as a bonus.
This has been discussed at Plumbers last year and the agreement was to
create VDSO plain (no namespace support) and VDSO extra (namespace
support).
The performance hit of the conditional + one etra cache line for tasks
which do not use a time name space is measurable.
That also avoids the whole mess of dealing with the static branch being
flipped. The time name space property is determined on fork/exec _before_
anything can touch the VDSO and depending on it the approriate version of
the VDSO is loaded.
Thanks,
tglx
