[vlc-devel] [PATCH] core: Add atomic refcounter helper

[Rémi Denis-Courmont]

Le sunnuntaina 1. heinäkuuta 2018, 22.10.20 EEST Romain Vimont a écrit :
> On Sun, Jul 01, 2018 at 09:48:03PM +0300, Rémi Denis-Courmont wrote:
> > Le sunnuntaina 1. heinäkuuta 2018, 20.56.03 EEST Romain Vimont a écrit :
> > > On Sun, Jul 01, 2018 at 08:10:54PM +0300, Rémi Denis-Courmont wrote:
> > > > Le sunnuntaina 1. heinäkuuta 2018, 12.30.33 EEST Romain Vimont a écrit 
:
> > > > > On Sun, Jul 01, 2018 at 11:59:08AM +0300, Rémi Denis-Courmont wrote:
> > > > > > Furthermore the boolean evaluation of the return value is the
> > > > > > opposite
> > > > > > of
> > > > > > what any other ref counting API does (other than VLC's). Usually,
> > > > > > zero
> > > > > > is
> > > > > > for last reference and non-zero for non-last references. Again,
> > > > > > this
> > > > > > is
> > > > > > prone to errors.
> > > > > 
> > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/t
> > > > > ree/
> > > > > lib/ refcount.c?h=v4.17#n208
> > > > 
> > > > This is very new.
> > > 
> > > In 2009, there were already atomic_dec_and_test(), with the same
> > > behavior:
> > > <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree
> > > /inc lude/asm-generic/atomic.h?h=v2.6.31#n123>
> > 
> > So what?
> 
> So Linux have, at least since 2009, a function very similar to the
> vlc_atomic_rc_dec() I propose, with the same behavior, and used in the
> very same way. Here are random examples in 2009 (there are many others):
> <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/inc
> lude/net/ax25.h?h=v2.6.31#n171>
> <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/in
> clude/net/netlabel.h?h=v2.6.31#n252>
> This contrasts with:
> > the boolean evaluation of the return value is the opposite of what
> > any other ref counting API does (other than VLC's). Usually, zero is
> > for last reference and non-zero for non-last references.

Not really. Using a boolean is confusing because different APIs have it 
different ways. By comparison, it is utterly obvious what atomic_fetch_OP 
(value before op) or atomic_OP_return (value after op).

> and:
> > This is very new.

Absolutely not. The refcount API in kernel is much newer than 2009. In fact, 
it is so new that it is not even in AOSP yet. And it is largelly by security 
people trying to remove a class of security issues that is only relevant to 
the kernel.

-- 
レミ・デニ-クールモン
http://www.remlab.net/