Re: [PATCH v2] socket: Provide put_cmsg_whitelist() for constant size copies

[Kees Cook]

On Tue, Feb 6, 2018 at 2:03 AM, David Miller <davem@xxxxxxxxxxxxx> wrote:
> From: Kees Cook <keescook@xxxxxxxxxxxx>
> Date: Fri, 2 Feb 2018 02:27:49 -0800
>
>> @@ -343,6 +343,14 @@ struct ucred {
>>
>>  extern int move_addr_to_kernel(void __user *uaddr, int ulen, struct sockaddr_storage *kaddr);
>>  extern int put_cmsg(struct msghdr*, int level, int type, int len, void *data);
>> +/*
>> + * Provide a bounce buffer for copying cmsg data to userspace when the
>> + * target memory isn't already whitelisted for hardened usercopy.
>> + */
>> +#define put_cmsg_whitelist(_msg, _level, _type, _ptr) ({             \
>> +             typeof(*(_ptr)) _val = *(_ptr);                         \
>> +             put_cmsg(_msg, _level, _type, sizeof(_val), &_val);     \
>> +     })
>
> I understand what you are trying to achieve, but it's at a real cost
> here.  Some of these objects are structures, for example the struct
> sock_extended_err is 16 bytes.

It didn't look like put_cmsg() was on a fast path, so it seemed like a
bounce buffer was the best solution here (and it's not without
precedent).

> And now we're going to copy it twice, once into the on-stack copy,
> and then once again into the CMSG blob.
>
> Please find a way to make hardened user copy happy without adding
> new overhead.

Another idea would be breaking put_cmsg() up into a macro with helper
functions, rearrange the arguments to avoid the math, and leaving the
copy_to_user() inline to see the const-ness, but that seemed way
uglier to me.

I'll think about it some more, but I think having put_cmsg_whitelist()
called only in a few places is reasonable here.

-Kees

-- 
Kees Cook
Pixel Security