[MODERATED] Re: spectrev1+

[Jon Masters <jcm@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2143 bytes --]

On 06/04/2018 03:51 PM, speck for Josh Poimboeuf wrote:
> On Mon, Jun 04, 2018 at 09:10:50PM +0200, speck for Peter Zijlstra wrote:
>> On Mon, Jun 04, 2018 at 12:36:23PM -0500, speck for Josh Poimboeuf wrote:
>>> On Mon, Jun 04, 2018 at 10:18:09AM -0700, speck for Dave Hansen wrote:
>>>> On 06/04/2018 10:14 AM, speck for Josh Poimboeuf wrote:
>>>>> So is the general idea that we're going to insert 3,000 lfences???  Or
>>>>> is there some reasonable way to weed out false positives?
>>>>
>>>> I think the original ~800 "hits" from Coverity ended up in something
>>>> like 10 places that we had to mitigate, either with lfence or
>>>> array_index_nospec().  They had to be culled down manually, though.
>>>
>>> Right, but since the Smatch v1 scanner was introduced in April, that has
>>> ballooned to 50 array_index_nospec() call sites.  In many (most?) cases
>>> I think it's difficult to definitively say "this is a false positive"
>>> with full certainty.
>>>
>>> So I'm just wondering what (if any) criteria we're using to make that
>>> determination.
>>
>> Yeah, that Coverity stuff missed a metric ton of sites.
> 
> Based on what Jiri said, it sounds like *three* metric tons.
> 
>> Many of the smatch results need fixing, mostly because even if you
>> cannot find a cache side-channel after most of them, there's no saying
>> there's no other side-channel hiding in plain sight.
>>
>> And if you kill all possible user input speculation, you don't have to
>> worry about someone changing the code and adding whatever it needs
>> later.
> 
> So we need to add ~3,000 array_index_nospec() calls, preferably before
> next Tuesday.  No problem!

Given the severity of this (sub)variant (which allows more precise
manipulation of program control flow during speculation) I really think
next Tuesday is not the time to be having the unembargo. I made a
request once again during a call with Intel for this to be delayed.

There's another MS patch Tuesday next month folks...I'll go ask
Microsoft to ask Intel for the same.

Jon.

-- 
Computer Architect | Sent from my Fedora powered laptop