[openib-general] [PATCH v4 2/7] IB/ipath - Implement new verbs DMA mapping functions

[Or Gerlitz]

Ralph Campbell wrote:
> On Tue, 2006-12-12 at 10:51 +0200, Or Gerlitz wrote:
>> Roland Dreier wrote:
>>>  > I would like to see this last set of patches integrated as is.
>>>  > I would like to get more experience with the current implementation
>>>  > before extending it to support other configurations.
>>>
>>> Yeah, let's go with that.  Since ipath depends on 64BIT in Kconfig
>>> anyway I think this is OK for now.
>> This design of ib_dma_map_single, ib_sg_dma_address etc returning u64 
>> instead of dma_addr_t causes the resulted patch to the IB ULPs to be 
>> quite big.
> 
> I think it was you who pointed out that dma_addr_t is
> 32 bits on sparc64.  Did you have a different solution
> in mind?

To be precise, I have pointed on a problem and you have come with the 
solution of having ib_dma_map_xxx work with u64 instead of dma_addr_t.

As Roland suggested, you could implement SW IOTLB that works with 
dma_addr_t and you have chosen not to.

>> Have you tested any dma_map single (eg IPoIB) and sg (eg SRP or iSER) 
>> consumer with this code?

> Yes.

The new API (eg ib_dma_map_xxx, ib_sg_dma_address and ib_sa_dma_len) 
adds some branching on each call, I wonder if you have seen any 
performance difference before/after the change. Specifically with IPoIB 
running a test with many PPS (ie iperf udp) or SRP IOPS test ?

Or.