Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 18 Nov 2009 08:41:54 -0500
From:      John Baldwin <jhb@freebsd.org>
To:        Jilles Tjoelker <jilles@stack.nl>
Cc:        svn-src-head@freebsd.org, Dmitry Pryanishnikov <lynx.ripe@gmail.com>, svn-src-all@freebsd.org, src-committers@freebsd.org, Edwin Groothuis <edwin@freebsd.org>
Subject:   Re: svn commit: r194783 - head/lib/libc/stdtime
Message-ID:  <200911180841.55183.jhb@freebsd.org>
In-Reply-To: <20091117182501.GA70742@stack.nl>
References:  <4B01E548.7040708@gmail.com> <20091117182501.GA70742@stack.nl>

next in thread | previous in thread | raw e-mail | index | archive | help
On Tuesday 17 November 2009 1:25:01 pm Jilles Tjoelker wrote:
> On Tue, Nov 17, 2009 at 01:50:32AM +0200, Dmitry Pryanishnikov wrote:
> > > Author: edwin
> > > Date: Tue Jun 23 22:28:44 2009
> > > New Revision: 194783
> > > URL: http://svn.freebsd.org/changeset/base/194783
> 
> > > Log:
> > >   Remove duplicate if-statement on gmt_is_set in gmtsub().
> 
> > >   MFC after:	1 week
> 
> > > Modified:
> > >   head/lib/libc/stdtime/localtime.c
> 
> >    This change looks like a (small?) pessimization to me: before it, 
> > _MUTEX_LOCK/_MUTEX_UNLOCK pair would be skipped for the case gmt_is_set 
> > == TRUE (all invocations except the first one), now it won't. I'm not 
> > sure whether this is critical here though...
> 
> It is certainly less efficient, but the old code was (most likely)
> wrong. It used an idiom known as "double checked locking", which is
> incorrect in most memory models. The problem is that the store to
> gmt_is_set may become visible without stores to other memory (gmtptr and
> what it points to) becoming visible.

That is easily fixed with a memory barrier.  Just use atomic_store_rel() to 
set gmt_is_set at the end of the setup phase.

-- 
John Baldwin



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200911180841.55183.jhb>