Re: data dependency naming inconsistency

From: Akira Yokosawa
Date: Thu Oct 14 2021 - 18:48:18 EST


On Thu, 14 Oct 2021 23:29:43 +0900, Akira Yokosawa wrote:
> [-CC akys: my 2nd address]
> On Thu, 14 Oct 2021 01:37:17 -0400, Michael S. Tsirkin wrote:
>> On Thu, Oct 14, 2021 at 01:43:24PM +0900, Akira Yokosawa wrote:
>>> On Mon, 11 Oct 2021 07:07:08 -0400, Michael S. Tsirkin wrote:
>>>> Hello Paul, all!
>>>
>>> Hello Michael,
>>>
>>> I thought Paul would respond soon, but looks like he has not
>>> done so.

This is because Michael used Paul's old email address.

Forwarding to his current address.

Paul, you can see the thread at the lore archive:
https://lore.kernel.org/lkml/20211011064233-mutt-send-email-mst@xxxxxxxxxx/T/

Thanks, Akira

>>> So, I'm trying to give some hint to your findings.
>>>
>>>> I've been reading with interest Paul's posts about Rust interactions with LKMM
>>>> https://paulmck.livejournal.com/63316.html
>>>> and in particular it states:
>>>> A data dependency involves a load whose return value directly or
>>>> indirectly determine the value stored by a later store, which results in
>>>> the load being ordered before the store.
>>>>
>>>> This matches the perf book:
>>>> A data dependency occurs when the value returned by
>>>> a load instruction is used to compute the data stored by
>>>> a later store instruction.
>>>
>>> You might likely be aware, but these concern "data dependency",
>>> not a _barrier_.
>>>
>>>>
>>>> however, memory-barriers.txt states:
>>>>
>>>> A data dependency barrier is a partial ordering on interdependent loads
>>>> only; it is not required to have any effect on stores, independent loads
>>>> or overlapping loads.
>>>>
>>>> It also says:
>>>> A data-dependency barrier is not required to order dependent writes
>>>> because the CPUs that the Linux kernel supports don't do writes
>>>> until they are certain (1) that the write will actually happen, (2)
>>>> of the location of the write, and (3) of the value to be written.
>>>
>>> These concern the historic "data-dependency barrier", or
>>> [smp_]read_barrier_depends(), which existed until Linux kernel v4.14.
>
> Ah... I should have said ", which existed prior to Linux kernel v4.15".
> This invited off-by-one error below...
>
>>>
>>>>
>>>> so the result it the same: writes are ordered without a barrier,
>>>> reads are ordered by a barrier.
>>>>
>>>> However, it would seem that a bit more consistency in naming won't
>>>> hurt.
>>>
>>> So, I don't think the historic term of "data-dependency barrier"
>>> can be changed.
>>>
>>> I guess the right approach would be to further de-emphasize
>>> "data-dependency barrier"/"data dependency barrier" in
>>> memory-barriers.txt.
>>>
>>> Rewrite by commit 8ca924aeb4f2 ("Documentation/barriers: Remove
>>> references to [smp_]read_barrier_depends()") did some of such
>>> changes, but it failed to update the introductory section of
>>> "VARIETIES OF MEMORY BARRIER".
>>> The part Michael quoted above belongs to it.
>>> I don't think it has any merit keeping it around.
>>>
>>> Also, there remain a couple of ascii-art diagrams concerning
>>> <data dependency barrier> in the first part of "EXAMPLES OF MEMORY
>>> BARRIER SEQUENCES" section, which, I think, can be removed as well.
>>>
>>> Hope this helps clarify the circumstances.
>>
>> It does, thanks! It might be worth adding a sentence along the lines of
>>
>> "NB: a data dependency barrier is distinct from a data dependency: it's
>> a barrier that used to be required in the presence of a data dependency.
>> Since v4.14 Linux no longer offers an API for a data dependency barrier.
>
> Since v4.15
>
>> Instead, using READ_ONCE is sufficient for ordering in the presence of a
>> data dependency".
>
>
> Maybe.
>
> But I'm more inclined to get rid of remaining contents related to the
> "data dependency barrier".
>
> Thanks, Akira
>
>>
>>
>>> Paul, what is your take on the naming of "data dependency"/
>>> "data dependency barrier"?
>>>
>>> Thanks, Akira
>>>
>>>>
>>>> Thanks,
>>>>
>>>> --
>>>> MST
>>