Re: KASAN+netlink, was: [PATCH] [net-next?] hns: avoid stack overflow with CONFIG_KASAN

2017-02-08 16:10 GMT+03:00 Arnd Bergmann [off-list ref]:

On Wed, Feb 8, 2017 at 1:24 PM, Johannes Berg [off-list ref] wrote:

quoted

Btw, what's causing this to start with? Can't the compiler reuse the
stack places?

I have no idea. It's trying to find out of bounds accesses for
objects on the stack, so maybe it gives each variable a separate
stack location in order to see which one caused problems?

If compiler cannot prove that access to the local variable is valid it
will add redzones around that variable
to be able to detect out of bounds accesses.

For example:
    static inline int nla_put_u8(struct sk_buff *skb, int attrtype, u8 value)
    {
           return nla_put(skb, attrtype, sizeof(u8), &value);
    }
compiler will surround 'value' with redzones to catch potential oob
access in nla_put().


Another way to fix this, would be something like this:

#ifdef CONFIG_KASAN
/* don't bloat stack */
#define __noinline_for_kasan    __noinline __maybe_unused
#else
#define __noinline_for_kasan    inline
#endif

static __noinline_for_kasan int nla_put_u8(struct sk_buff *skb, int
attrtype, u8 value)
{
       return nla_put(skb, attrtype, sizeof(u8), &value);
}