lkml.org 
[lkml]   [2020]   [Oct]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Bug in herd7 [Was: Re: Litmus test for question from Al Viro]
From
Date
On Sat, 3 Oct 2020 13:13:38 -0400, Alan Stern wrote:
> On Sun, Oct 04, 2020 at 12:16:31AM +0900, Akira Yokosawa wrote:
>> Hi Alan,
>>
>> Just a minor nit in the litmus test.
>>
>> On Sat, 3 Oct 2020 09:22:12 -0400, Alan Stern wrote:
>>> To expand on my statement about the LKMM's weakness regarding control
>>> constructs, here is a litmus test to illustrate the issue. You might
>>> want to add this to one of the archives.
>>>
>>> Alan
>>>
>>> C crypto-control-data
>>> (*
>>> * LB plus crypto-control-data plus data
>>> *
>>> * Expected result: allowed
>>> *
>>> * This is an example of OOTA and we would like it to be forbidden.
>>> * The WRITE_ONCE in P0 is both data-dependent and (at the hardware level)
>>> * control-dependent on the preceding READ_ONCE. But the dependencies are
>>> * hidden by the form of the conditional control construct, hence the
>>> * name "crypto-control-data". The memory model doesn't recognize them.
>>> *)
>>>
>>> {}
>>>
>>> P0(int *x, int *y)
>>> {
>>> int r1;
>>>
>>> r1 = 1;
>>> if (READ_ONCE(*x) == 0)
>>> r1 = 0;
>>> WRITE_ONCE(*y, r1);
>>> }
>>>
>>> P1(int *x, int *y)
>>> {
>>> WRITE_ONCE(*x, READ_ONCE(*y));
>>
>> Looks like this one-liner doesn't provide data-dependency of y -> x on herd7.
>
> You're right. This is definitely a bug in herd7.
>
> Luc, were you aware of this?
>
>> When I changed P1 to
>>
>> P1(int *x, int *y)
>> {
>> int r1;
>>
>> r1 = READ_ONCE(*y);
>> WRITE_ONCE(*x, r1);
>> }
>>
>> and replaced the WRITE_ONCE() in P0 with smp_store_release(),
>> I got the result of:
>>
>> -----
>> Test crypto-control-data Allowed
>> States 1
>> 0:r1=0;
>> No
>> Witnesses
>> Positive: 0 Negative: 3
>> Condition exists (0:r1=1)
>> Observation crypto-control-data Never 0 3
>> Time crypto-control-data 0.01
>> Hash=9b9aebbaf945dad8183d2be0ccb88e11
>> -----
>>
>> Restoring the WRITE_ONCE() in P0, I got the result of:
>>
>> -----
>> Test crypto-control-data Allowed
>> States 2
>> 0:r1=0;
>> 0:r1=1;
>> Ok
>> Witnesses
>> Positive: 1 Negative: 4
>> Condition exists (0:r1=1)
>> Observation crypto-control-data Sometimes 1 4
>> Time crypto-control-data 0.01
>> Hash=843eaa4974cec0efae79ce3cb73a1278
>> -----
>
> What you should have done was put smp_store_release in P0 and left P1 in
> its original form. That test should not be allowed, but herd7 says that
> it is.

Yea, that was what I tried first, expecting the result of "Never".

>
>> As this is the same as the expected result, I suppose you have missed another
>> limitation of herd7 + LKMM.
>
> It would be more accurate to say that we all missed it. :-) (And it's
> a bug in herd7, not a limitation of either herd7 or LKMM.) How did you
> notice it?

:-) :-) :-)

Well, I thought I had never seen a litmus test with such one-liner.
So I split the READ_ONCE() and WRITE_ONCE() into two lines and
got the expected result.

I don't expect much from herd7's C mode in the first place.
(No offense intended!)


>> By the way, I think this weakness on control dependency + data dependency
>> deserves an entry in tools/memory-model/Documentation/litmus-tests.txt.
>>
>> In the LIMITATIONS section, item #1 mentions some situation where
>> LKMM may not recognize possible losses of control-dependencies by
>> compiler optimizations.
>>
>> What this litmus test demonstrates is a different class of mismatch.
>
> Yes, one in which LKMM does not recognize a genuine dependency because
> it can't tell that some optimizations are not valid.
>
> This flaw is fundamental to the way herd7 works. It examines only one
> execution at a time, and it doesn't consider the code in a conditional
> branch while it's examining an execution where that branch wasn't taken.
> Therefore it has no way to know that the code in the unexecuted branch
> would prevent a certain optimization. But the compiler does consider
> all the code in all branches when deciding what optimizations to apply.

I see.

>
> Here's another trivial example:
>
> r1 = READ_ONCE(*x);
> if (r1 == 0)
> smp_mb();
> WRITE_ONCE(*y, 1);
>
> The compiler can't move the WRITE_ONCE before the READ_ONCE or the "if"
> statement, because it's not allowed to move shared memory accesses past
> a memory barrier -- even if that memory barrier isn't always executed.
> Therefore the WRITE_ONCE actually is ordered after the READ_ONCE, but
> the memory model doesn't realize it.>
>> Alan, can you come up with an update in this regard?
>
> I'll write something.

Thanks!

Akira

>
> Alan
>

\
 
 \ /
  Last update: 2020-10-04 00:51    [W:0.161 / U:24.360 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site