Re: p = kmalloc(sizeof(*p), )

From: Linus Torvalds <torvalds@osdl.org>
To: Al Viro <viro@ftp.linux.org.uk>
Cc: Willy Tarreau <willy@w.ods.org>, Robert Love <rml@novell.com>,
	Russell King <rmk+lkml@arm.linux.org.uk>,
	Linux Kernel List <linux-kernel@vger.kernel.org>
Subject: Re: p = kmalloc(sizeof(*p), )
Date: Sun, 18 Sep 2005 10:31:36 -0700 (PDT)	[thread overview]
Message-ID: <Pine.LNX.4.58.0509181028140.26803@g5.osdl.org> (raw)
In-Reply-To: <20050918171845.GL19626@ftp.linux.org.uk>

On Sun, 18 Sep 2005, Al Viro wrote:
> 
> That's why you do
> 	*p = (struct foo){....};
> instead of
> 	memset(p, 0, sizeof...);
> 	p->... =...;

Actually, some day that migth be a good idea, but at least historically, 
gcc has really really messed that kind of code up.

Last I looked, depending on what the initializer was, gcc would create a 
temporary struct on the stack first, and then do a "memcpy()" of the 
result. Not only does that obviously generate a lot of extra code, it also 
blows your kernel stack to kingdom come.

So be careful out there, and check what code it generates first. With at 
least a few versions of gcc.

(For _small_ structures it's wonderful. As far as I can tell, gcc does a
pretty good job on structs that are just a single long-word in size).

		Linus