Date: Wed, 18 Jun 2014 07:25:11 +0200 (CEST)
From: Julia Lawall <julia.lawall@lip6.fr>
To: Joe Perches <joe@perches.com>
cc: Jesper Juhl <jj@chaosbits.net>, Fabian Frederick <fabf@skynet.be>,
        linux-kernel@vger.kernel.org, Gilles Muller <Gilles.Muller@lip6.fr>
Subject: Re: [PATCH 1/1] scripts/coccinelle/free: add conditional kfree
 test
In-Reply-To: <1403042299.2649.3.camel@joe-AO725>
Message-ID: <alpine.DEB.2.02.1406180711210.2059@localhost6.localdomain6>
References: <1403034199-9824-1-git-send-email-fabf@skynet.be>  <1403035539.2722.3.camel@joe-AO725>  <alpine.DEB.2.02.1406172331001.2055@localhost6.localdomain6> <1403042299.2649.3.camel@joe-AO725>
User-Agent: Alpine 2.02 (DEB 1266 2009-07-14)
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org


On Tue, 17 Jun 2014, Joe Perches wrote:

> (adding Jesper Juhl)
> 
> On Tue, 2014-06-17 at 23:33 +0200, Julia Lawall wrote:
> > On Tue, 17 Jun 2014, Joe Perches wrote:
> > > On Tue, 2014-06-17 at 21:43 +0200, Fabian Frederick wrote:
> > > > This patch adds a trivial script warning on
> > > > 
> > > > if(foo)
> > > > 	kfree(foo)
> > > > 
> > > > (based on checkpatch)
> > > []
> > > > diff --git a/scripts/coccinelle/free/cond_kfree.cocci b/scripts/coccinelle/free/cond_kfree.cocci
> > > []
> > > > +* if (E)
> > > > +*	kfree@p(E);
> > > 
> > > You should probably add all of the unnecessary
> > > conditional tests that checkpatch uses:
> > > 
> > > kfree
> > > usb_free_urb
> > > debugfs_remove
> > > debugfs_remove_recursive
> > 
> > Personally, I would prefer that the message encourage the user to consider 
> > whether it is necessary to call these functions with NULL as an argument 
> > in any case.
> 
> Jesper quite awhile ago wrote:
> 
> https://lkml.org/lkml/2005/10/13/81
> 
> - Since kfree always checks for a NULL argument there's no reason to have an
> additional check prior to calling kfree. It's redundant.
> - In many cases gcc produce significantly smaller code without the redundant
> check before the call.
> - It's been shown in the past (in discussions on LKML) that it's generally a
> win performance wise to avoid the extra NULL check even though it might save
> a function call. Only when the NULL check avoids the function call in the vast
> majority of cases and the code is in a hot path does it make sense to have it.
> - This patch removes quite a few more source lines than it adds, cutting down
> on the overall number of source lines is generally a good thing.
> - This patch reduces the indentation level, which is nice when the kfree call
> is inside some deeply nested construct.

What I don't like is:

a = kmalloc(...);
if (!a) goto out;
b = kmalloc(...);
if (!b) goto out;
c = kmalloc(...);
if (!c) goto out;
...
out:
  kfree(a);
  kfree(b);
  kfree(c);

With a little thought, one could reorganize the code to not call kfree on 
a null value.  Someone who is not familiar with Linux programming style 
could interpret the feedback as that the above code is perfectly fine.  
(And perhaps some people do consider that it is perfectly fine).

On the other hand, in the case:

x = NULL;
if (complicated_condition)
  x = kmalloc(...);
  if (!x) return;
y = something(...);
if (!y)
  goto out1;
...
out1: kfree(x);

I guess it's OK.  Mildly unpleasant, but probably the best option given 
the various tradeoff.

In looking at Jesper's patch, I see that another case is:

a = kmalloc(...);
b = kmalloc(...);
if (!a || !b) {
  kfree(a);
  kfree(b);
}

Personally, I would rather see each call have its own error handling code.  
There is no point to make the second call if the first one fails.

When one tries to understand code, the main questions are why is this done 
here, and why is this not done here.  Doing things that are unnecessary 
introduces confusion in this regard.  Perhaps it doesn't matter for 
kmalloc and kfree because everyone is familiar with them and they are 
pretty innocuous.  But for the more obscure functions, like in my 
recollection of Markus's patch, I'm not convinced that simply blindly 
removing all unneeded tests without thinking whether the code could be 
written in a better way is a good idea.

julia
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/