Re: dynamic-sg patch has broken rdma_rxe

[Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>]

On 20/10/2020 13:56, Jason Gunthorpe wrote:
> On Tue, Oct 20, 2020 at 01:31:27PM +0100, Tvrtko Ursulin wrote:
>>
>> On 20/10/2020 12:47, Jason Gunthorpe wrote:
>>> On Tue, Oct 20, 2020 at 12:37:05PM +0100, Tvrtko Ursulin wrote:
>>>
>>>>> Why put this confusing code in every caller? Especially for something
>>>>> a driver is supposed to call. Will just make bugs
>>>>
>>>> For max_segment to be aligned is a requirement today so callers are
>>>> ready.
>>>
>>> No, it turns out all the RDMA drivers were became broken when they
>>> converted to use the proper U32_MAX for their DMA max_segment size,
>>> then they couldn't form SGLs anymore.
>>>
>>> I don't want to see nonsense code like this:
>>>
>>>           dma_set_max_seg_size(dev->dev, min_t(unsigned int, U32_MAX & PAGE_MASK,
>>>                                                SCATTERLIST_MAX_SEGMENT));
>>>
>>> In drivers.
>>>
>>> dma_set_max_seg_size is the *hardware* capability, and mixing in
>>> things like PAG_MASK here is just nonsense.
>>
>> Code was obviously a no-op non-sense.
>>
>> So the crux of the argument is that U32_MAX is conceptually the right thing
>> which defines the DMA max_segment size? Not some DMA define or anything, but
>> really U32_MAX? And that all/some DMA hardware does not think in pages but
>> really in bytes?
> 
> Yes.
> 
> The HW has 32 bits for a length field, so U32_MAX accurately defines
> the HW DMA capability

Not just the max, but the granularity as well. If byte granularity (well 
less than PAGE_SIZE/4k) exists in some hw then I agree the API change 
makes sense.

Regards,

Tvrtko