Date: Tue, 17 May 2011 12:56:40 -0400 From: John Baldwin <jhb@freebsd.org> To: Max Laier <max@love2party.net> Cc: neel@freebsd.org, Andriy Gapon <avg@freebsd.org>, Attilio Rao <attilio@freebsd.org>, FreeBSD current <freebsd-current@freebsd.org>, Stephan Uphoff <ups@freebsd.org>, Peter Grehan <grehan@freebsd.org> Subject: Re: proposed smp_rendezvous change Message-ID: <201105171256.41091.jhb@freebsd.org> In-Reply-To: <4DD2A058.6050400@love2party.net> References: <4DCD357D.6000109@FreeBSD.org> <4DD26720.3000001@FreeBSD.org> <4DD2A058.6050400@love2party.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tuesday, May 17, 2011 12:20:40 pm Max Laier wrote:
> On 05/17/2011 05:16 AM, John Baldwin wrote:
> ...
> > Index: kern/kern_switch.c
> > ===================================================================
> > --- kern/kern_switch.c (revision 221536)
> > +++ kern/kern_switch.c (working copy)
> > @@ -192,15 +192,22 @@
> > critical_exit(void)
> > {
> > struct thread *td;
> > - int flags;
> > + int flags, owepreempt;
> >
> > td = curthread;
> > KASSERT(td->td_critnest != 0,
> > ("critical_exit: td_critnest == 0"));
> >
> > if (td->td_critnest == 1) {
> > + owepreempt = td->td_owepreempt;
> > + td->td_owepreempt = 0;
> > + /*
> > + * XXX: Should move compiler_memory_barrier() from
> > + * rmlock to a header.
> > + */
>
> XXX: If we get an interrupt at this point and td_owepreempt was zero,
> the new interrupt will re-set it, because td_critnest is still non-zero.
>
> So we still end up with a thread that is leaking an owepreempt *and*
> lose a preemption.
I don't see how this can still leak owepreempt. The nested interrupt should
do nothing (except for possibly set owepreempt) until td_critnest is 0.
However, we can certainly lose preemptions.
I wonder if we can abuse the high bit of td_critnest for the owepreempt flag
so it is all stored in one cookie. We only set owepreempt while holding
thread_lock() (so interrupts are disabled), so I think we would be ok and not
need atomic ops.
Hmm, actually, the top-half code would have to use atomic ops. Nuts. Let me
think some more.
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201105171256.41091.jhb>