On Fri, Jul 24, 2015 at 01:48:09PM -0500, Steve Wise wrote: > > The use of FRWR for RDMA READ should be iWarp specific, IB shouldn't > > pay that overhead. I am expecting to see a cap_rdma_read_rkey or > > something in here ? > > Ok. But cap_rdma_read_rkey() doesn't really describe the > requirement. The requirement is rkey + REMOTE_WRITE. So it is more > like rdma_cap_read_requires_remote_write() which is ugly and too > long (but descriptive)... I don't care much what name you pick, just jam something like this in the description If set then RDMA_READ must be performed by mapping the local buffers through a rkey MR with ACCESS_REMOTE_WRITE enabled. The rkey of this MR should be passed in as the sg_lists's lkey for IB_WR_RDMA_READ_WITH_INV. FRWR should be used to register the buffer in the send queue, and the read should be issued using IB_WR_RDMA_READ_WITH_INV (xx can we just implicitly rely on this? Are there any iWarp cards that support FRWR but not WITH_INV?) Finally, only a single SGE can be used with RDMA_READ, all scattering must be accomplished with the MR. This quite dramatically changes what is an allowed scatter for the transfer, IB can support arbitary unaligned S/G lists, while this is now forced into gapless page aligned elements. Your patch takes care of this? And only impacts IB? Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
