[PATCHv3 5/6] ARM: imx: Add initial support for Freescale LS1021A

[Jingchang Lu]

>-----Original Message-----
>From: Arnd Bergmann [mailto:arnd at arndb.de]
>Sent: Wednesday, September 10, 2014 3:42 PM
>To: Lu Jingchang-B35083
>Cc: linux-arm-kernel at lists.infradead.org; Guo Shawn-R65073;
>devicetree at vger.kernel.org
>Subject: Re: [PATCHv3 5/6] ARM: imx: Add initial support for Freescale
>LS1021A
>
>On Wednesday 10 September 2014 03:31:19 Jingchang Lu wrote:
>
>> >> +DT_MACHINE_START(LS1021A, "Freescale LS1021A")  #ifdef
>> >> +CONFIG_ZONE_DMA
>> >> +       .dma_zone_size  = SZ_128M,
>> >> +#endif
>> >> +       .init_machine   = ls1021a_init_machine,
>> >> +       .dt_compat      = ls1021a_dt_compat,
>> >> +       .restart        = mxc_restart,
>> >> +MACHINE_END
>> >
>> >I believe someone recently posted a patch to derive the dma_zone_size
>> >from device tree. Can yo try to find that and see if that will work for
>you?
>> >
>> >Can you explain what the reason is for needing a DMA zone?
>>
>> With LPAE enabled on our SoC, we meet the system complaint of
>> "coherent DMA mask 0xffffffff is smaller than system GFP_DMA mask
>> 0xffffffffffffffff", and I notice that CONFIG_ZONE_DMA and dma_zone_size
>is a common resolve for this.
>
>Ok, I see. The actual point of dma_zone_size however is slightly different,
>and I think you should not use it like this. We normally only use ZONE_DMA
>if there are devices that have a limitation smaller than 4GB, and that
>appear to be the case for your system.
>
>The message you quote is only present in arch/powerpc, so I'm not sure
>what symptoms you are actually seeing. Please try removing the
>dma_zone_size setting for your platform and report if it works or what the
>symptom is if it does not work with the latest kernel.
>
>We definitely need to get this to work out of the box without a
>dma_zone_size hack.
>
>Can you describe what the memory layout is of your platform? Can you have
>RAM installed above the 4GB physical address boundary? If you can, are
>there any devices that are unable to perform DMA into that memory without
>the use of an IOMMU?
>
>	Arnd
Our first LS1021A support is based on kernel-3.12, where when LPAE enabled
it will compare the device's coherent_dma_mask with arm_dma_limit, which is
64-bit 0xffffffffffffffff without CONFIG_ZONE_DMA, with CONFIG_ZONE_DMA can
limit the comparison and avoid the warning.
All device can address 32-bit address space on LS1021A, With you comment above,
I remove the dma_zone_size and only reserve the CONFIG_ZONE_DMA On kernel-3.12,
finding all works well, so I will remove the dma_zone_size setting on 3.12. Thanks.

I have a look on latest kernel, finding get_coherent_dma_mask() has limited the mask
to (u64)DMA_BIT_MASK(32), does this mean the CONFIG_ZONE_DMA is not needed for me
here? Thanks.

Best Regards,
Jingchang