Re: [PATCH] fix calls to request_module()

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 10 Dec 2008 21:03:37 -0800 Roland Dreier <rdreier@cisco.com> wrote:

> The kernel has such code eg in init/main.c, which does
> 
> 	printk(linux_banner);
> 
> when linux_banner is only visible to the compiler as
> 
> extern const char linux_banner[];
> 
> however the trivial fix
> 
> diff --git a/init/main.c b/init/main.c
> index 7e117a2..e471598 100644
> --- a/init/main.c
> +++ b/init/main.c
> @@ -568,7 +568,7 @@ asmlinkage void __init start_kernel(void)
>  	boot_cpu_init();
>  	page_address_init();
>  	printk(KERN_NOTICE);
> -	printk(linux_banner);
> +	printk("%s", linux_banner);
>  	setup_arch(&command_line);
>  	mm_init_owner(&init_mm, &init_task);
>  	setup_command_line(command_line);
> 
> doesn't seem that appealing, since it bloats the object code for a
> non-bug -- 7 bytes for me on x86_64:
> 
> add/remove: 0/0 grow/shrink: 1/0 up/down: 7/0 (7)
> function                                     old     new   delta
> start_kernel                                 680     687      +7
> 
> given the number of such warnings I see in a typical compile, this would
> be a fairly hefty amount of bloat just to shut up gcc.

yes, that would suck.  otoh, our current warning spew actually causes bugs.

I wonder if we could add a printk_stfu() which isn't declared
attribute(printf) and which simply calls printk.  We might still get a
single warning at the interface point.

> On the other hand, gcc warning on such code (untrusted format string
> passed into a printf-like function) seems quite legitimate as well.

Yes, we've had actual bugs in the kernel from this, where the control
string was user-provided.  root-only user, fortunately.