Re: [PATCH V2] mmc: sdhci-acpi: Add 64-bit DMA support

[Adrian Hunter <adrian.hunter@intel.com>]

On 28/10/14 12:18, Arnd Bergmann wrote:
> On Tuesday 28 October 2014 12:05:30 Adrian Hunter wrote:
>> On 28/10/14 11:43, Arnd Bergmann wrote:
>>> On Tuesday 28 October 2014 10:37:20 Adrian Hunter wrote:
>>>>  static int sdhci_acpi_enable_dma(struct sdhci_host *host)
>>>>  {
>>>> -       return 0;
>>>> +       struct sdhci_acpi_host *c = sdhci_priv(host);
>>>> +       struct device *dev = &c->pdev->dev;
>>>> +       int err = -1;
>>>> +
>>>> +       if (c->dma_setup)
>>>> +               return 0;
>>>> +
>>>> +       if (host->flags & SDHCI_USE_64_BIT_DMA) {
>>>> +               if (host->quirks2 & SDHCI_QUIRK2_BROKEN_64_BIT_DMA) {
>>>> +                       host->flags &= ~SDHCI_USE_64_BIT_DMA;
>>>> +               } else {
>>>> +                       err = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64));
>>>> +                       if (err)
>>>> +                               dev_warn(dev, "Failed to set 64-bit DMA mask\n");
>>>> +               }
>>>> +       }
>>>> +
>>>> +       if (err)
>>>> +               err = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32));
>>>> +
>>>> +       c->dma_setup = !err;
>>>> +
>>>> +       return err;
>>>>  }
>>>>
>>>
>>> I don't think it's worth a dev_warn() message (maybe dev_info), there is nothing
>>
>> It is worth a dev_warn because 32-bit DMA can allocate memory for bounce
>> buffers which jeopardizes memory reclaim.
> 
> Then you should also warn if SDHCI_USE_64_BIT_DMA isn't or if
> SDHCI_QUIRK2_BROKEN_64_BIT_DMA is set I guess.

The warning is for when the controller supports 64-bit, not when it doesn't.

> 
>>> wrong in this case, but you probably have to clear the SDHCI_USE_64_BIT_DMA to
>>> ensure that the driver won't try to to use the 64-bit DMA if the mask is not set.
>>
>> 64-bit DMA will work with a 32-bit DMA mask so there is no need.
> 
> Wouldn't that at least be (slightly) less efficient? You end up having to do
> two expensive MMIO register accesses each time you transfer an address.

I doubt it is measureable and I don't want to optimize a degenerate case anyway.