Re: [RFC net-next 05/10] net/smc: register RMB-related memory region

[Ursula Braun <ubraun-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

On 07/20/2017 05:17 PM, Sagi Grimberg wrote:
>> A memory region created for a new RMB must be registered explicitly,
>> before the peer can make use of it for remote DMA transfer.
> 
> OK, Now I understand the sgl construction in the previous patch.
> 
> We got rid of the synchronous memory registration API in the kernel
> in the absence of users, but now with smc we might want to restore
> it because looking at the patchset, the fast memory registration
> scheme is obviously not a good fit.
> 
> For the simple smc use case the API should be as simple as:
> 
> ib_reg_contig_mr(struct ib_pd *pd, void *addr, size_t length);
> 
> Do you have a use-case for scattered buffers memory registration?
> 

>From my smc perspective I would like to finally allocate the RMB memory as 
scattered pages mapped to contiguous DMA-memory. When creating a new connection
the smc protocol exchanges some information, among them the dma-address to
enable the peer for RDMA WRITEs. According to the smc protocol just one dma-address
can be exchanged; thus contiguous DMA-memory describable by one address only is
required here.
I have prepared already such an smc patch, and it worked for s390, but on my
x-system the scattered pages mapped to scattered DMA-memory after calling
ib_dma_map_sg() and ib_map_mr_sg(). Have I missed something? Is there a way
to explicitly ask for mapping to contiguous memory? 

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html