On 11/10/2015 1:25 PM, Jason Gunthorpe wrote:
On Tue, Nov 10, 2015 at 04:04:32AM -0800, Christoph Hellwig wrote:
On Tue, Nov 10, 2015 at 01:46:40PM +0200, Sagi Grimberg wrote:
On 10/11/2015 13:41, Christoph Hellwig wrote:
Oh, and while we're at it. Can someone explain why we're even
using rdma_read_chunk_frmr for IB? It seems to work around the
fact tat iWarp only allow a single RDMA READ SGE, but it's used
whenever the device has IB_DEVICE_MEM_MGT_EXTENSIONS, which seems
wrong.
I think Steve can answer it better than I can. I think that it is
just to have a single code path for both IB and iWARP. I agree that
the condition seems wrong and for small transfers rdma_read_chunk_frmr
is really a performance loss.
Well, the code path already exists, but only is used fi
IB_DEVICE_MEM_MGT_EXTENSIONS isn't set. Below is an untested patch
that demonstrates how I think svcrdma should setup the reads. Note
that this also allows to entirely remove it's allphys MR.
Note that as a followon this would also allow to remove the
non-READ_W_INV code path from rdma_read_chunk_frmr as a future
step.
I like this, my only comment is we should have a rdma_cap for this
behavior, rdma_cap_needs_rdma_read_mr(pd) or something?
Windows NDKPI has this, it's the oh-so-succinctly-named flag
NDK_ADAPTER_FLAG_RDMA_READ_SINK_NOT_REQUIRED. The ULP is free
to ignore it and pass the NDK_MR_FLAG_RDMA_READ_SINK flag anyway,
the provider is expected to ignore it if not needed.
+ if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num)) {
Use here
+ /*
+ * iWARP requires remote write access for the data sink, and
+ * only supports a single SGE for RDMA_READ requests, so we'll
+ * have to use a memory registration for each RDMA_READ.
+ */
+ if (!(dev->device_cap_flags &
IB_DEVICE_MEM_MGT_EXTENSIONS)) {
Lets enforce this in the core, if rdma_cap_needs_rdma_read_mr is set
the the device must also set IB_DEVICE_MEM_MGT_EXTENSIONS, check at
device creation time.
+ } else if (rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num)) {
+ /*
+ * For IB or RoCE life is easy, no unsafe write access is
+ * required and multiple SGEs are supported, so we don't need
+ * to use MRs.
+ */
+ newxprt->sc_reader = rdma_read_chunk_lcl;
+ } else {
+ /*
+ * Neither iWarp nor IB-ish, we're out of luck.
+ */
goto errout;
No need for the else, !rdma_cap_needs_rdma_read_mr means pd->local_dma_lkey is okay
to use.
Hmm, agreed, but it must still be acceptable to perform a registration
instead of using the local_dma_lkey. As mentioned earlier, there are
scatter limits and other implications for all-physical addressing that
an upper layer may choose to avoid.
Tom.
--
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
