Re: [PATCH 1/2] compiler.h: Introduce ptr_eq() to preserve address dependency

[Alan Huang <mmpgouride@gmail.com>]

2024年9月29日 07:55，Boqun Feng <boqun.feng@gmail.com> wrote：
> 
> 
> 
> On Sun, Sep 29, 2024, at 6:26 AM, Alan Huang wrote:
>> 2024年9月28日 23:55，Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote：
>>> 
>>> On 2024-09-28 17:49, Alan Stern wrote:
>>>> On Sat, Sep 28, 2024 at 11:32:18AM -0400, Mathieu Desnoyers wrote:
>>>>> On 2024-09-28 16:49, Alan Stern wrote:
>>>>>> On Sat, Sep 28, 2024 at 09:51:27AM -0400, Mathieu Desnoyers wrote:
>>>>>>> equality, which does not preserve address dependencies and allows the
>>>>>>> following misordering speculations:
>>>>>>> 
>>>>>>> - If @b is a constant, the compiler can issue the loads which depend
>>>>>>>   on @a before loading @a.
>>>>>>> - If @b is a register populated by a prior load, weakly-ordered
>>>>>>>   CPUs can speculate loads which depend on @a before loading @a.
>>>>>> 
>>>>>> It shouldn't matter whether @a and @b are constants, registers, or
>>>>>> anything else.  All that matters is that the compiler uses the wrong
>>>>>> one, which allows weakly ordered CPUs to speculate loads you wouldn't
>>>>>> expect it to, based on the source code alone.
>>>>> 
>>>>> I only partially agree here.
>>>>> 
>>>>> On weakly-ordered architectures, indeed we don't care whether the
>>>>> issue is caused by the compiler reordering the code (constant)
>>>>> or the CPU speculating the load (registers).
>>>>> 
>>>>> However, on strongly-ordered architectures, AFAIU, only the constant
>>>>> case is problematic (compiler reordering the dependent load), because
>>>> I thought you were trying to prevent the compiler from using one pointer
>>>> instead of the other, not trying to prevent it from reordering anything.
>>>> Isn't this the point the documentation wants to get across when it says
>>>> that comparing pointers can be dangerous?
>>> 
>>> The motivation for introducing ptr_eq() is indeed because the
>>> compiler barrier is not sufficient to prevent the compiler from
>>> using one pointer instead of the other.
>> 
>> barrier_data(&b) prevents that.
>> 
> 
> It prevents that because it acts as barrier() + OPTIMIZER_HIDE_VAR(b).
> I don’t see much value of using that since we can resolve the problem
> with OPTIMIZER_HIDE_VAR() alone.

Yeah, barrier_data generates more instructions.

> 
> Regards,
> Boqun
> 
>>> 
>>> But it turns out that ptr_eq() is also a good tool to prevent the
>>> compiler from reordering loads in case where the comparison is
>>> done against a constant.
>>> 
>>>>> CPU speculating the loads across the control dependency is not an
>>>>> issue.
>>>>> 
>>>>> So am I tempted to keep examples that clearly state whether
>>>>> the issue is caused by compiler reordering instructions, or by
>>>>> CPU speculation.
>>>> Isn't it true that on strongly ordered CPUs, a compiler barrier is
>>>> sufficient to prevent the rcu_dereference() problem?  So the whole idea
>>>> behind ptr_eq() is that it prevents the problem on all CPUs.
>>> 
>>> Correct. But given that we have ptr_eq(), it's good to show how it
>>> equally prevents the compiler from reordering address-dependent loads
>>> (comparison with constant) *and* prevents the compiler from using
>>> one pointer rather than the other (comparison between two non-constant
>>> pointers) which affects speculation on weakly-ordered CPUs.
>>> 
>>>> You can make your examples as specific as you like, but the fact remains
>>>> that ptr_eq() is meant to prevent situations where both:
>>>> The compiler uses the wrong pointer for a load, and
>>>> The CPU performs the load earlier than you want.
>>>> If either one of those doesn't hold then the problem won't arise.
>>> 
>>> Correct.
>>> 
>>> Thanks,
>>> 
>>> Mathieu
>>> 
>>> 
>>> -- 
>>> Mathieu Desnoyers
>>> EfficiOS Inc.
>>> https://www.efficios.com