Re: [RFC PATCH for 4.18 00/14] Restartable Sequences

[Daniel Colascione]

On Wed, May 2, 2018 at 1:23 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Wed, May 02, 2018 at 06:27:22PM +0000, Daniel Colascione wrote:
> > On Wed, May 2, 2018 at 10:22 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx>
wrote:
> > >> On Wed, May 02, 2018 at 03:53:47AM +0000, Daniel Colascione wrote:
> > > > Suppose we make a userspace mutex implemented with a lock word
having
> > three
> > > > bits: acquired, sleep_mode, and wait_pending, with the rest of the
word
> > not
> > > > being relevant at the moment.
> >
> > > So ideally we'd kill FUTEX_WAIT/FUTEX_WAKE for mutexes entirely, and
go
> > > with FUTEX_LOCK/FUTEX_UNLOCK that have the same semantics as the
> > > existing FUTEX_LOCK_PI/FUTEX_UNLOCK_PI, namely, the word contains the
> > > owner TID.
> >
> > That doesn't work if you want to use the rest of the word for something
> > else, like a recursion count. With FUTEX_WAIT and FUTEX_WAKE, you can
make
> > a lock with two bits.

> Recursive locks are teh most horrible crap ever. And having the tid in

What happened to providing mechanism, not policy?

You can't wish away recursive locking. It's baked into Java and the CLR,
and it's enshrined in POSIX. It's not going away, and there's no reason not
to support it efficiently.

> the word allows things like kernel based optimistic spins and possibly
> PI related things.

Sure. A lot of people don't want PI though, or at least they want to opt
into it. And we shouldn't require an entry into the kernel for what we can
in principle do efficiently in userspace.

> > > As brought up in the last time we talked about spin loops, why do we
> > > care if the spin loop is in userspace or not? Aside from the whole PTI
> > > thing, the syscall cost was around 150 cycle or so, while a LOCK
CMPXCHG
> > > is around 20 cycles. So ~7 spins gets you the cost of entry.
> >
> > That's pre-KPTI, isn't it?

> Yes, and once the hardware gets sorted, we'll be there again. I don't
> think we should design interfaces for 'broken' hardware.

It would be a mistake to design interfaces under the assumption that
everyone has fast permission level transitions.