Re: [RFC] mmc: cqhci: commit descriptors before setting the doorbell

From: Faiz Abbas
Date: Fri Oct 18 2019 - 03:22:11 EST


Adrian,

On 16/10/19 5:46 PM, Faiz Abbas wrote:
> Adrian,
>
> On 15/10/19 7:15 PM, Adrian Hunter wrote:
>> On 15/10/19 10:55 AM, Faiz Abbas wrote:
>>> Hi,
>>>
>>> On 15/10/19 12:08 AM, Faiz Abbas wrote:
>>>> Add a write memory barrier to make sure that descriptors are actually
>>>> written to memory before ringing the doorbell.
>>>>
>>>> Signed-off-by: Faiz Abbas <faiz_abbas@xxxxxx>
>>>> ---
>>>>
>>>> This patch fixes a very infrequent ADMA error (1 out of 100 times) that
>>>> I have been seeing after enabling command queuing for J721e.
>>>> Also looking at memory-barriers.txt and this commit[1],
>>>> it looks like we should be doing this before any descriptor write
>>>> followed by a doorbell ring operation. It'll be nice if someone with more
>>>> expertise in memory barriers can comment.
>>>>
>>>> [1] ad1a1b9cd67a ("scsi: ufs: commit descriptors before setting the
>>>> doorbell")
>>>
>>> So I see that cqhci_readl/writel() use readl/writel_relaxed() which
>>> seems to be causing this issue. Should I just fix this by converting
>>> those to readl/writel with memory barriers instead?
>>
>> Perhaps we could do both changes i.e. add wmb() and convert to non-relaxed
>> readl/writel
>>
>
> readl is implemented as readl_relaxed(); __rmb();
> and
> writel is implemented as wmb(); writel_relaxed();
>
> I think another wmb() before writel will be redundant.
>
> Maybe this patch is good enough in itself.
>

Do you agree?

Thanks,
Faiz