ask for help about swiotlb buffer is full

[wangkefeng.wang@huawei.com (Kefeng Wang)]

On 2015-02-06 2:19, Catalin Marinas wrote:
> On Wed, Feb 04, 2015 at 02:48:03PM +0000, Ding Tianhong wrote:
>> On 2015/2/4 21:32, Catalin Marinas wrote:
>>> On Wed, Feb 04, 2015 at 12:01:01PM +0000, Kefeng Wang wrote:
>>>> On 2015-02-03 2:24, Catalin Marinas wrote:
>>>>> On Sat, Jan 31, 2015 at 04:00:06AM +0000, Kefeng Wang wrote:
>>>>>> 4kb page,  use Make ARCH=arm64 defconfig in v3.19 to generate config.
>>>>>>
>>>>>> [    0.678293] software IO TLB [mem 0x7e800000-0x7ec00000] (4MB) mapped at [ffffffc07e800000-ffffffc07ebfffff]
>>>>>> [    0.686991] DMA: preallocated 256 KiB pool for atomic allocations
>>>>>
>>>>> Was the swiotlb buffer size the same in the 3.16 kernel? The only thing
>>>>> I recall adding was the atomic pool allocations but these are only for
>>>>> non-coherent DMA ops and only for dma_alloc/free. I assume, in the case
>>>>> of SATA, the failure is on the dma_map_sg() path.
>>>>
>>>> Swiotlb buffer size is both 4M in v3.16 and v3.19-rc4, and the failure is
>>>> on the dma_map_sg() -> swiotlb_map_sg_attrs.
>>>>
>>>>> Maybe with a 3.19 kernel you get more than 4MB swiotlb buffers used at a
>>>>> time with your tests; can you try increasing this via a kernel command
>>>>> like to, let's say, 8MB? If I got my calculations correctly (an IO TLB
>>>>> slab is 1 << 11):
>>>>>
>>>>> 	swiotlb=4096
>>>>>
>>>>> If it still runs out with bigger buffers, we may need to look into
>>>>> potential leaks.
>>>>
>>>> The buddy allocator can only support 4M contiguous physical memory, so it's
>>>> useless to increase swiotlb buffer.
>>>
>>> You could hack arch/arm64/Kconfig to set a higher
>>> CONFIG_FORCE_MAX_ZONEORDER as a test. Depending on the test result, we
>>> can look for an alternative solution.
>>
>> I have try this before and could fix the problem, but I think it is
>> not a perfect solution,.
> 
> So it's not some bug leaking memory but a genuine need for bigger
> swiotlb buffer. Ideally your hardware should have an iommu as bouncing
> is not cheap.
> 
> Most architectures using swiotlb seem to use the default size of 64MB
> (with the risk of wasting too much memory on smaller systems). That's
> what we had on arm64 before commit 3690951fc6d42f3a (arm64: Use swiotlb
> late initialisation) but the problem was that the generic swiotlb_init()
> function didn't bother with which zone it allocated memory from and it
> wasn't always suitable for 32-bit DMA.

Thanks for you help, I have tested this patch, and sata could work fine again.
Will you merger this patch?


> 
> Below is an attempt to move back to early swiotlb initialisation but
> fixing the memblock low memory allocation to make it suitable for 32-bit
> only devices.
>