Date: Mon, 26 Jan 2009 09:36:17 -0500 From: John Baldwin <jhb@freebsd.org> To: freebsd-arch@freebsd.org Cc: arch@freebsd.org, Alfred Perlstein <alfred@freebsd.org>, Julian Elischer <julian@elischer.org>, Kip Macy <kmacy@freebsd.org> Subject: Re: need for another mutex type/flag? Message-ID: <200901260936.18232.jhb@freebsd.org> In-Reply-To: <20090126105140.GL5889@elvis.mu.org> References: <497BA91D.805@elischer.org> <497D5DF8.8000706@elischer.org> <20090126105140.GL5889@elvis.mu.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Monday 26 January 2009 5:51:40 am Alfred Perlstein wrote:
> * Julian Elischer <julian@elischer.org> [090125 22:53] wrote:
> > Alfred Perlstein wrote:
> > >
> > >Jeff, I think that Julian really wants to prevent a sleep inside
> > >his context. Right now, I think we only check for mutexes held
> > >before a sleep that arne't sleepable. It might make sense to allow
> > >one to just mark a thread non-sleepable even though no mutex is
> > >held.
> > >
> > >Julian, is that right?
> >
> > basically, though I don't know the details of implementation..
> > I just know that mutexes per se aren't bad for netgraph but
> > that node authors need some guidance on how to use them and
> > some way to prove to them when they do the wrong thing.
>
> The way to add the assertion you want would be to keep a count
> inside of the thread structure "td_nosleep", set to 0 at thread
> creation, then you can do this:
>
> TD_SLEEP_NO(td); /* td->td_nosleep++ */
> call_some_untrusted_code();
> TD_SLEEP_OK(td); /* td->td_nosleep-- */
>
> Then add this to subr_witness.c:witness_warn():
>
> if (flags & WARN_SLEEPOK && td->td_nosleep != 0) {
> printf("Sleeping in unsleepable context.\n");
> n++; /* this variable is local to witness_warn()
> and triggers an ASSERT at the end */
> }
>
> I could have sworn we already had such a feature, but it appears
> that it's only accessable if you're holding a lock, if you added
> this counter, then you could catch sleeps without needing a lock
> held.
We have this feature already for sleeping, but I think Julian isn't worried
about sleeping (i.e. *sleep() or cv_*wait*()), but wants to prevent the code
from acquiring any other locks. It's easy to add a MTX_LEAF, I'm just not
sure if we really want to micro-manage the code that much. WITNESS will
already catch any LORs, and if they are acquiring a rarely-contested lock
then they aren't going to back up traffic in the common case.
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200901260936.18232.jhb>