Re: [PATCH] mmc: sdhci-xenon: Fix 2G limitation on AC5 SoC

[Adrian Hunter]

On 9/11/22 10:40, Vadym Kochan wrote:
> Hi Adrian,
> 
> On Wed, 9 Nov 2022 09:50:06 +0200, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>> On 8/11/22 21:05, Vadym Kochan wrote:
>>> Hi Adrian,
>>>
>>> On Thu, 13 Oct 2022 09:40:00 +0300, Vadym Kochan <vadym.kochan@xxxxxxxxxxx> wrote:
>>>> Hi Robin,
>>>>
>>>> On Mon, 22 Aug 2022 11:06:43 +0100, Robin Murphy <robin.murphy@xxxxxxx> wrote:
>>>>> On 2022-08-21 07:17, Christoph Hellwig wrote:
>>>>>> On Thu, Aug 18, 2022 at 03:07:40PM +0300, Vadym Kochan wrote:
>>>>>>> It works with the following changes:
>>>>>>>
>>>>>>>      #1 dma-ranges = <0x0 0x0 0x2 0x0 0x0 0x80000000>;
>>>>>>>
>>>>>>>      #3 swiotlb="force"
>>>>>>>
>>>>>>> Is it OK to force the memory allocation from the start for the swiotlb ?
>>>>>>
>>>>>> It should be ok, but isn't really optimal.
>>>>>>
>>>>>> I wonder if we should just allow DT to specify the swiotlb buffer
>>>>>> location.  Basically have yet another RESERVEDMEM_OF_DECLARE variant
>>>>>> for it, which shouldn't be all that much work except for figuring
>>>>>> out the interaction with the various kernel command line options.
>>>>>
>>>>> We already have all the information we need in the DT (and ACPI), the 
>>>>> arm64 init code just needs to do a better job of interpreting it 
>>>>> properly. I'll see what I can come up with once I've finished what I'm 
>>>>> currently tied up in.
>>>>>
>>>>> Thanks,
>>>>> Robin.
>>>>
>>>> Sorry to disturb you, I just 'd like to know if you have
>>>> some ideas to share or patches to test ?
>>>>
>>>> Thank you!
>>>>
>>>
>>> Since AC5X eMMC controller can fail to work on boards with >2GB memory,
>>> and considering that the best fix may not be easy (as it requires arm64 infra changes),
>>> so would it be OK to use PIO mode as temporary solution ?
>>>
>>> I understand that arm64 changes might not be trivial and it might take significant
>>> amount of time to implement considering this unusual case, I just think that better
>>> to make eMMC working even if it will be slow.
>>
>> You can disable DMA if you wish:
>> 	SDHCI_QUIRK_BROKEN_DMA | SDHCI_QUIRK_BROKEN_ADMA
>> however did you consider SDMA:
>> 	SDHCI_QUIRK_BROKEN_ADMA
>> which uses a bounce buffer allocated by SDHCI?
>>
>> In any case, you need to add comments to the code
>> and commit message explaining the swiotlb issue.
>>
> 
> There is a snip from my earlier reply:
> 
> [snip]
> 
>> I could use DMA only in 2 ways:
>>
>> #1 Use sdhci bounce buffer with SDMA mode
>>
>>     But there was the issue that SDMA requires that SDHCI v4 mode should
>>     be enabled, and when I enable it via sdhci_enable_v4_mode(host)
>>     then I got error that EXT_CSD can't be recognized.
>>
>>     But if I comment this line in sdhci.c:
>>     
>>     int sdhci_setup_host(struct sdhci_host *host)
>>     {
>>         ...
>>
>> 	/* SDMA does not support 64-bit DMA if v4 mode not set */
>> 	if ((host->flags & SDHCI_USE_64_BIT_DMA) && !host->v4_mode) {

Do you need 64-bit DMA?  The 2G limitation is still
within 32-bits isn't it?

>> 		pr_info("XXX SDMA does not support 64-bit DMA if v4 mode not set\n");
>> 		host->flags &= ~SDHCI_USE_SDMA;
>> 	}
>>
>>        ...
>>     }
>>
>>     then everything is OK.
>>
>> #2 Use restricted-dma-pool in device-tree
>>
>>    But I am not sure if it is good solution compared to #1.
>>
>> Setting only DMA mask did not help because after some time I got
>> "DMA overflow address" error stack-traces.
> 
> [snip]
> 
> Regards,
> Vadym