Re: Thread-safe libio

[Nathan Myers <ncm@best.com>]

Jeff Anton <jaanton@jps.net> wrote:
> Ulrich Drepper wrote:
> > Andreas Gruenbacher <a.gruenbacher@infosys.tuwien.ac.at> writes:
> > 
> > >  a) use flockfile() / funlockfile() to lock the file for each I/O
> > > operation.
> > >     Advantage: simple
> > >     Disadvantage: high overhead
> > 
> > It is no question that this is the method to be used.
>
> There is a lot of reason _not_ to do this.  If this kills performance
> then the whole of the iostream library will be shot.  To my mind this
> option is completely unacceptable.

I agree with Jeff.  

It would be madness to kill iostream performance, slowing it by two 
orders of magnitude, unnecessarily.  The Standard iostream facilities 
define a way to get locking, and the Standard operators << and >> will
use it.

The approach Andreas suggested would penalize standard-conforming
code for no net benefit to non-standard code.  The Standard-blessed 
way to get thread-safety in streams is to place a counting mutex to 
be claimed by the [io]stream::sentry constructor, and decremented or
released by its destructor.  These operations would normally occur 
much less frequently than per-character.

Nathan Myers
ncm@cantrip.org