Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 7 Jul 2017 09:16:17 +0100
From: Ard Biesheuvel <>
To: Arnd Bergmann <>
Cc: Kernel Hardening <>, Kees Cook <>, 
	Linus Torvalds <>
Subject: Re: [RFC/RFT PATCH] gcc-plugins: force initialize auto variables
 whose addresses are taken

On 6 July 2017 at 23:08, Arnd Bergmann <> wrote:
> On Thu, Jul 6, 2017 at 1:25 PM, Arnd Bergmann <> wrote:
>> On Thu, Jul 6, 2017 at 1:09 PM, Arnd Bergmann <> wrote:
>>> On Thu, Jul 6, 2017 at 12:13 PM, Ard Biesheuvel
>> Sorry, bad example, that one is a bit less undefined than
>> I thought, as it will produce the same result every time,
>> regardless of the stack contents. I'll try to come up
>> with another test program instead.
> I've tried a few more things, but couldn't actually come up with an example
> that ends up using uninitialized stack values without also warning about it,
> so your plugin may actually cover the most important cases.
> The remaining cases I found are either uninitialized uses that we get
> a compile-time warning for, or other kinds of undefined behavior
> (as in my earlier example).

Yes, so your original example is actually the opposite side of the
same coin. The compiler notices that 'i' may be returned
uninitialized, but since 0 is equally suitable as any other value in
this case, it just makes the assignment unconditional.

So imagine a function call h(&i) preceding the switch. This forces the
compiler to return whatever value i happens to have, because it cannot
know whether calling h() amounts to an initialization.

This all comes back to the way we pass structs by reference only,
which is why I retained the struct/union type test in the plugin, even
though stack buffers (i.e, u8 buf[xxx]) would probably deserve the
same treatment.

I think we should be able to deal with this more elegantly by
annotating function parameters that are guaranteed to be set by the
function, e.g,

extern h(int * __out i)

where we would on the one hand enforce that every code path in the
definition of h() contains an assignment of *i, and on the other hand,
omit the forced initialization of variables that appear as __out

Alternatively (or to complement the effort), we could at least start
making the effort to convert byref initializers to struct assignment,
in cases where the function is a void type to begin with.

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.