Re: [PATCH 1/2] [v4] net: emac: emac gigabit ethernet controller driver
From: Rob Herring
Date: Fri Apr 15 2016 - 08:36:23 EST
On Thu, Apr 14, 2016 at 6:34 PM, Timur Tabi <timur@xxxxxxxxxxxxxx> wrote:
> Vikram Sethi wrote:
>>>>
>>>> >> retval = dma_coerce_mask_and_coherent(&pdev->dev,
>>>> >> DMA_BIT_MASK(64));
>>>> >> if (retval) {
>>>> >> dev_err(&pdev->dev, "failed to set DMA mask err %d\n",
>>>> >> retval);
>>>> >> goto err_res;
>>>> >> }
>>
>> How can you set the mask to 64 bits when the EMAC IP on FSM9900 and
>> QDF2432 can only do 32 bit DMA?
>> The mask in that API is a bit mask describing which bits of an address
>> your device supports.
>
>
> Vikram, Shanker, and I discussed this offline, and came to a consensus.
>
> The FSM9900 is a 32-bit platform, so the kernel will never create a DMA
> address above 4GB. Even if the driver sets the mask to 64 bits, it will
> technically work. However, the mask should be set to 32 because all address
> buses are 32 bits.
>
> The QDF2432 is different. Although it's an ARM64 platform, we have the
> unfortunate situation that only 32 bits of that address is wired to the rest
> of the chip. So even though the Emac can handle 64-bit bus addresses, if it
> actually attempts to DMA above 4GB, the address will get truncated and
> corrupt memory. The mask needs to be set to 32.
>
> There may or may not be other ARM64 chips from us that won't have this
> problem in the future, so these hypothetical chips would have a mask of 64.
>
> So I think the solution is to create a device tree (and ACPI) property that
> holds the mask.
>
> dma-mask = <0 0xffffffff>;
>
> or
>
> dma-mask = <0xffffffff 0xffffffff>;
No. See dma-ranges.
Rob
>
> The driver will then do this:
>
> u64 dma_mask;
> device_property_read_u64(&pdev->dev, "dma-mask", &dma_mask);
> dma_coerce_mask_and_coherent(&pdev->dev, dma_mask);
>
> What I'm not sure yet is whether I should call
> dma_coerce_mask_and_coherent() or dma_set_coherent_mask().
>
> --
> Qualcomm Innovation Center, Inc.
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> Forum, a Linux Foundation collaborative project.