Re: [PATCH 4/4] irqchip: bcm2836: Use a more generic memory barrier call

[Eric Anholt <eric@anholt.net>]

[-- Attachment #1: Type: text/plain, Size: 1816 bytes --]

Stephen Warren <swarren@wwwdotorg.org> writes:

> On 04/08/2016 12:20 PM, Eric Anholt wrote:
>> Stephen Warren <swarren@wwwdotorg.org> writes:
>>
>>> On 04/04/2016 09:44 PM, Eric Anholt wrote:
>>>> dsb() requires an argument on arm64, so we needed to add "sy".
>>>> Instead, take this opportunity to switch to the same smp_wmb() call
>>>> that gic uses for its IPIs.  This is a less strong barrier than we
>>>> were doing before (dmb(ishst) compared to dsb(sy)), but it seems to be
>>>> the correct one.
>>>
>>> I assume all MMIO is part of the ish domain?
>>>
>>> If so, the series,
>>> Acked-by: Stephen Warren <swarren@wwwdotorg.org>
>>
>> I don't know if this barrier implies ordering all the way out to AXI on
>> this HW, but I don't think that's a requirement of this function.
>
> My understanding was that the barrier was explicitly to work around a 
> bug in the bus fabric of the SoC, and hence the barrier very much does 
> have to affect the transaction all the way out to AXI. Re-reading 
> BCM2835-ARM-Peripherals.pdf section 1.3 "Peripheral access precautions 
> for correct memory ordering" seems to confirm this.

My understanding of the explicit barrier here, which was copied from
other irqchips, is "Make sure that normal memory writes before our IPI
on this CPU appear on the other CPUs before they get the IPI" (like the
comment says).  This barrier was not put in to deal with the
283x-specific weird AXI behavior.

Note that we had previously decided that the weird AXI ordering
behavior, which is about repeated reads or repeated writes from the same
CPU across different peripherals, is already covered by the barriers
present in readl() and writel().  The writel() barrier happens to be a
dsb() as well, so this explicit barrier is actually redundant.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 818 bytes --]