Re: [PATCH] net: core: Fix Spectre v1 vulnerability

["Gustavo A. R. Silva" <gustavo@embeddedor.com>]

On 12/22/18 9:00 PM, Alexei Starovoitov wrote:
> On Sat, Dec 22, 2018 at 08:53:40PM -0600, Gustavo A. R. Silva wrote:
>> Hi,
>>
>> On 12/22/18 8:40 PM, David Miller wrote:
>>> From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
>>> Date: Sat, 22 Dec 2018 15:59:54 -0800
>>>
>>>> On Sat, Dec 22, 2018 at 03:07:22PM -0800, David Miller wrote:
>>>>> From: "Gustavo A. R. Silva" <gustavo@embeddedor.com>
>>>>> Date: Fri, 21 Dec 2018 14:49:01 -0600
>>>>>
>>>>>> flen is indirectly controlled by user-space, hence leading to
>>>>>> a potential exploitation of the Spectre variant 1 vulnerability.
>>>>>>
>>>>>> This issue was detected with the help of Smatch:
>>>>>>
>>>>>> net/core/filter.c:1101 bpf_check_classic() warn: potential spectre issue 'filter' [w]
>>>>>>
>>>>>> Fix this by sanitizing flen before using it to index filter at line 1101:
>>>>>>
>>>>>> 	switch (filter[flen - 1].code) {
>>>>>>
>>>>>> and through pc at line 1040:
>>>>>> 	
>>>>>> 	const struct sock_filter *ftest = &filter[pc];
>>>>>>
>>>>>> Notice that given that speculation windows are large, the policy is
>>>>>> to kill the speculation on the first load and not worry if it can be
>>>>>> completed with a dependent load/store [1].
>>>>>>
>>>>>> [1] https://marc.info/?l=linux-kernel&m=152449131114778&w=2
>>>>>>
>>>>>> Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
>>>>>
>>>>> BPF folks, I'll take this directly.
>>>>>
>>>>> Applied and queued up for -stable, thanks.
>>>>
>>>> hmm. what was the rush?
>>>> I think it is unnecessary change.
>>>> Though fp is passed initially from user space
>>>> it's copied into kernel struct first.
>>>> There is no way user space can force kernel to mispredict
>>>> branch in for (pc = 0; pc < flen; pc++) loop.
>> The following piece of code is the one that can be mispredicted, not the for
>> loop:
>>
>> 1013         if (flen == 0 || flen > BPF_MAXINSNS)
>> 1014                 return false;
>>
>> Instead of calling array_index_nospec() inside bpf_check_basics_ok(), I
>> decided to place the call close to the code that could be compromised. This
>> is when accessing filter[].
> 
> Why do you think it can be mispredicted?
>

Beause fprog->len comes from user space:

bpf_prog_create_from_user() -> bpf_check_basics_ok()

> I've looked at your other patch for nfc_sock_create() where you're adding:
> + proto = array_index_nospec(proto, NFC_SOCKPROTO_MAX);
> 
> 'proto' is the value passed in _register_ into system call.
> There is no need to wrap it with array_index_nospec().
> It's not a load from memory and user space cannot make it slow.
> Slow load is a necessary attribute to trigger speculative execution
> into mispredicted branch.
> 

We might be interpreting the information available about Spectre a bit 
different, but when the Spectre paper talks about memory vs cpu speed it 
seems to me that it's just an example to illustrate how the microcode 
can come into the equation and speculate. So I'm genuinely curious about 
your last statement: "Slow load is a necessary attribute..." Where did 
you get that info from?

Can't we have the case in which the code can be "trained" to read
perfectly valid values for prog->len for quite a while, making the
microcode come into place and speculate about:

1013         if (flen == 0 || flen > BPF_MAXINSNS)
1014                 return false;

and then make flen to be greater than BPF_MAXINSNS?

> What tool did you use to analyze this?
> Did you analyze all warnings case by case or just trusted the tool
> and generated these patches?
> 

I read every case, but I sometimes might be wrong of course.

Thanks
--
Gustavo