[Xen-devel] [PATCH] xen: Add __GFP_DMA flag when xen_swiotlb_init gets free pages.

[Roger Pau Monné]

El 21/04/15 a les 12.36, Stefano Stabellini ha escrit:
> On Tue, 21 Apr 2015, Ian Campbell wrote:
>> On Mon, 2015-04-20 at 18:54 +0100, Stefano Stabellini wrote:
>>> This should definitely be done only on ARM and ARM64, as on x86 PVH
>>> assumes the presence of an IOMMU. We need an ifdef.
>>>
>>> Also we need to figure out a way to try without GFP_DMA in case no ram
>>> under 4g is available at all, as some arm64 platforms don't have any. Of
>>> course in those cases we don't need to worry about devices and their dma
>>> masks.  Maybe we could use memblock for that?
>>
>> It's pretty ugly, but I've not got any better ideas.
>>
>> It would perhaps be less ugly as a an arch-specific
>> get_me_a_swiotlb_region type function, with the bare __get_free_pages as
>> the generic fallback.
> 
> We could do that, but even open code like this isn't too bad: it might
> be ugly but at least is very obvious.
> 
> 
>>>  Something like:
>>>
>>>     struct memblock_region *reg;
>>>     gfp_t flags = __GFP_NOWARN;
>>>
>>> #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
>>> 	for_each_memblock(memory, reg) {
>>> 		unsigned long start = memblock_region_memory_base_pfn(reg);
>>>
>>>         if (start < 4G) {
>>>             flags |= __GFP_DMA;
>>>             break;
>>>         }
>>>     }
>>> #endif
>>>
>>>     [...]
>>>
>>>     xen_io_tlb_start = (void *)__get_free_pages(flags, order);
>>>     
>>>
>>>
>>>
>>>
>>>> This is also conceptually wrong since it doesn't matter where the pages
>>>> are in PFN space,  but where they are in bus address (MFN) space (which
>>>> is what the subsequent hypercall is required to sort out).
>>>
>>> Actually on ARM dom0 is mapped 1:1, so it is the same thing.
>>
>> On a system with a fully functional SMMU dom0 may not be 1:1 mapped, but
>> I think that dom0 can still treat that as 1:1 mapped for these purposes,
>> since the SMMU will provide that illusion.
>>
>> Dumb question, and this might affect PVH too, if you have an IOMMU and a
>> device with a limited DMA range, I suppose you need to provide DMA
>> addresses in the <4G for the input to the IOMMU (i.e. PFN) and not the
>> output (i.e. MFN) space (since the device only sees PFNs).
> 
> I think you mean "for the input to the device (PFN)", but I presume the
> same.
> 
> 
>> So even for x86 PVH isn't something required here to ensure that the
>> swiotlb has suitable pages under 4GB in PFN space too?
>>
>> (On ARM PFN==IPA and MFN==PA)
> 
> I guess that is true. PVH people, any thoughts?

AFAIK Linux PVH uses the native swiotlb, and FreeBSD does the same. I
expect the device expects PFNs (not MFNs) below the 4GB range, or else
the design is broken because there's no way the guest can figure out if
the MFN behind a PFN is below 4GB.

Roger.