Re: [RFC][CFT][PATCHSET v1] uaccess unification

[Al Viro]

On Wed, Mar 29, 2017 at 04:43:05PM -0700, Linus Torvalds wrote:

> > As for __copy_in_user()... I'm not sure we want to keep it in the long run -
> 
> I agree, it's probably not worth it at all.
> 
> In fact, I suspect none of the "__copy_.*_user()" versions are worth
> it, and we should strive to remove them.
> 
> There aren't even that many users, and they _have_ caused security
> issues when people have had some path that hasn't checked the range.

Actually, looking through those users shows some very odd places: for example,
sctp_setsockopt_bindx() does
        /* Check the user passed a healthy pointer.  */
        if (unlikely(!access_ok(VERIFY_READ, addrs, addrs_size)))
                return -EFAULT;

        /* Alloc space for the address array in kernel memory.  */
        kaddrs = kmalloc(addrs_size, GFP_USER | __GFP_NOWARN);
        if (unlikely(!kaddrs))
                return -ENOMEM;

        if (__copy_from_user(kaddrs, addrs, addrs_size)) {
                kfree(kaddrs);
                return -EFAULT;
        }
The obvious question is "why not memdup_user()?" and rationale looks fishy:
 * We don't use copy_from_user() for optimization: we first do the
 * sanity checks (buffer size -fast- and access check-healthy
 * pointer); if all of those succeed, then we can alloc the memory
 * (expensive operation) needed to copy the data to kernel. Then we do
 * the copying without checking the user space area
 * (__copy_from_user()).
plus that:
    sctp: use GFP_USER for user-controlled kmalloc
    
    Dmitry Vyukov reported that the user could trigger a kernel warning by
    using a large len value for getsockopt SCTP_GET_LOCAL_ADDRS, as that
    value directly affects the value used as a kmalloc() parameter.
    
    This patch thus switches the allocation flags from all user-controllable
    kmalloc size to GFP_USER to put some more restrictions on it and also
    disables the warn, as they are not necessary.

First of all, access_ok() for sanity checks on size is BS - on some
architectures it's constant 1 and on *all* architectures it allows a lot
more than what kmalloc() will.  So it won't stop a malicious program from
getting to kmalloc() and wasting its cycles and it's not much help with
buggy ones.  __GFP_NOWARN part is more interesting, but... the quoted
commit misses memdup_user() calls in other setsockopt cases on the same
sctp.  So if we care about that one, we probably should care about the
rest of them as well, and I doubt that open-coding each is a good solution.
Something like kvmemdup_user(), perhaps?  I.e. quiet fallback to vmalloc
on large sizes/in case when kmalloc barfs, with kvfree on the freeing side?
Or just a flat-out check for some reasonably upper limit on optlen in
the very beginning?