On 2020/10/17 3:27, Florian Fainelli wrote: > On 10/16/20 11:23 AM, Arnd Bergmann wrote: >> On Fri, Oct 16, 2020 at 6:48 PM Florian Fainelli <f.fainelli@xxxxxxxxx> wrote: >>> On 10/16/20 4:01 AM, Arnd Bergmann wrote: >>>> On Fri, Oct 16, 2020 at 11:09 AM Zhen Lei <thunder.leizhen@xxxxxxxxxx> wrote: >>>>> >>>>> Suggested-by: Arnd Bergmann <arnd@xxxxxxxx> >>>>> Signed-off-by: Zhen Lei <thunder.leizhen@xxxxxxxxxx> >>>> >>>> Acked-by: Arnd Bergmann <arnd@xxxxxxxx> >>>> >>>> I see that at least the 'bcd' and 'xhci' devices in fact try to >>>> use 64-bit DMA. It would be good to test this on actual >>>> hardware to ensure that it works correctly when this is enabled. >>>> >>>> Ideally avoiding the swiotlb bounce buffering should only >>>> make it faster here, but there are many chips on which >>>> 64-bit DMA is broken in some form. >>> >>> Is this change really an improvement though? This 'usb' pseudo bus node >>> could just keep being defined with #address-cells = <1> and #size-cells >>> = <1> so as to satisfy the 'reg' definition however we could just adjust >>> dma-ranges to indicate full 64-bit addressing capability. Would not that >>> work? >> >> When #address-cells is '1', you cannot specify dma-ranges that >> go beyond a 32-bit address range. > > Would not it be enough to remove the 'dma-ranges' property though? Sorry > for being slow here. Remove the 'dma-ranges' property should also work. After all, it is equivalent to the original empty dma-ranges scheme. In addition, since the IOMMU nodes are defined, it should be enabled. Therefore, Arnd's concern about the scenario where the IOMMU is disabled may not be triggered. All roads lead to Rome. Which solution should be chose depends on individual preferences. For me, either solution is fine. The third solution is to define a non-empty dma-ranges property. However, because stingray-usb.dtsi is included in multiple files, it may not be appropriate. >
