Re: [PATCH 2/4] arch: Add lightweight memory barriers fast_rmb() and fast_wmb()

[Alexander Duyck <alexander.h.duyck@redhat.com>]

On 11/18/2014 03:58 AM, Will Deacon wrote:
> On Tue, Nov 18, 2014 at 03:13:29AM +0000, Alexander Duyck wrote:
>> On 11/17/2014 04:39 PM, Benjamin Herrenschmidt wrote:
>>> On Mon, 2014-11-17 at 12:24 -0800, Alexander Duyck wrote:
>>>> Yes and no.  So for example on ARM I used the dmb() operation, however
>>>> I
>>>> have to use the barrier at the system level instead of just the inner
>>>> shared domain.  However on many other architectures they are just the
>>>> same as the smp_* variants.
>>>>
>>>> Basically the resultant code is somewhere between the smp and non-smp
>>>> barriers in terms of what they cover.
>>> There I don't quite follow you. You need to explain better especially in
>>> the documentation because otherwise people will get it wrong...
>>>
>>> If it's ordering in the coherent domain, I fail to see how a DMA agent
>>> is different than another processor when it comes to barriers, so I fail
>>> to see the difference with smp_*
>>>
>>> I understand the MMIO vs. memory issue, we do have the same on powerpc,
>>> but that other aspect eludes me.
>>>
>> ARM adds some funky things.  They have two different types of
>> primitives, a dmb() which is a data memory barrier, and a dsb() which is
>> a data synchronization barrier.  Then with each of those they have the
>> "domains" the barriers are effective within.
>>
>> So for example on ARM a rmb() is dsb(sy) which means it is a system wide
>> synchronization barrier which stops execution on the CPU core until the
>> read completes.  However the smp_rmb() is a dmb(ish) which means it is
>> only a barrier as far as the inner shareable domain which I believe only
>> goes as far as the local shared cache hierarchy and only guarantees read
>> ordering without necessarily halting the CPU or stopping in-order
>> speculative reads.  So what a coherent_rmb() would be in my setup is
>> dmb(sy) which means the barrier runs all the way out to memory, and it
>> is allowed to speculative read as long as it does it in order.
>>
>> If it is still unclear you might check out Will Deacon's talk on the
>> topic at https://www.youtube.com/watch?v=6ORn6_35kKo, at about 7:00 in
>> he explains the whole domains thing, and at 13:30 he explains dmb()/dsb().
> So actually, this is an interesting case where the barrier would like to
> know whether the memory returned by dma_alloc_coherent is h/w coherent
> (normal, cacheable) or s/w coherent (normal, non-cacheable). I think Ben
> is thinking of the h/w coherent case (i.e. actual snooping into the CPU
> caches by the DMA master).
>
> For the former, we could use inner-shareable barriers. For the latter, we'd
> need to use outer-shareable barriers.
>
> If we can't tell, then these should be dmb(osh), which will work for both.
>
> Will

Okay, so I will update the ARM portion of my patches to use osh and 
oshst then since it sounds like I was using too strong of barriers.

- Alex