Re: [PATCH 16/18] net: mpls: prevent bounds-check bypass via speculative execution

[Dan Williams <dan.j.williams@intel.com>]

On Tue, Jan 9, 2018 at 5:57 PM, Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
> On Tue, Jan 09, 2018 at 04:48:24PM -0800, Dan Williams wrote:
>>
>> #define __nospec_array_ptr(base, idx, sz)                               \
>> ({                                                                      \
>>         union { typeof(&base[0]) _ptr; unsigned long _bit; } __u;       \
>>         unsigned long _i = (idx);                                       \
>>         unsigned long _s = (sz);                                        \
>>         unsigned long _v = (long)(_i | _s - 1 - _i)                     \
>>                                         >> BITS_PER_LONG - 1;           \
>>         unsigned long _mask = _v * ~0UL;                                 \
>>         OPTIMIZER_HIDE_VAR(_mask);                                      \
>>         __u._ptr = &base[_i & _mask];                                   \
>>         __u._bit &= _mask;                                              \
>>         __u._ptr;                                                       \
>> })
>
> _v * ~0UL doesn't seem right and non intuitive.
> What's wrong with:
>   unsigned long _mask = ~(long)(_i | _s - 1 - _i) >> BITS_PER_LONG - 1;

Yeah, I noticed it was ok immediately after I sent that.

> and why OPTIMIZER_HIDE_VAR ?

It was in Linus' original. but that was when it had the ternary
conditional, I'll drop it. It does not change the generated assembly.

> Could you remove '&' ?

Yes, that should be __u.ptr = base + (i & _mask)

> since in doesn't work for:
> struct {
>   int fd[4];
>   ...
> } *fdt;
> it cannot be used as array_acces(fdt->fd, ...);
>
> Could you please drop nospec_ prefix since it is misleading ?

When you came up with that tweak you noted:

"The following:
[..]
is generic and no speculative flows."

> This macro doesn't prevent speculation.

It masks dangerous speculation. At least, I read nospec as "No
Spectre" and it is a prefix used in the Spectre-v2 patches.

I also want to include the option, with a static branch, to switch it
to the hard "no speculation" version with an ifence if worse comes to
worse and we find a compiler / cpu where it doesn't work. The default
will be the fast and practical implementation.

> I think array_access() was the best name so far.

For other usages I need the pointer to the array element, also
array_access() by itself is unsuitable for __fcheck_files because we
still need rcu_dereference_raw() on the element de-reference. So, I
think it's better to get a sanitized array element pointer which can
be used with rcu, READ_ONCE(), etc... directly rather than try to do
the access in the same macro.