RDMA/rw: Fall back to direct SGE on MR pool exhaustion

When IOMMU passthrough mode is active, ib_dma_map_sgtable_attrs()
produces no coalescing: each scatterlist page maps 1:1 to a DMA
entry, so sgt.nents equals the raw page count. A 1 MB transfer
yields 256 DMA entries. If that count exceeds the device's
max_sgl_rd threshold (an optimization hint from mlx5 firmware),
rdma_rw_io_needs_mr() steers the operation into the MR
registration path. Each such operation consumes one or more MRs
from a pool sized at max_rdma_ctxs -- roughly one MR per
concurrent context. Under write-intensive workloads that issue
many concurrent RDMA READs, the pool is rapidly exhausted,
ib_mr_pool_get() returns NULL, and rdma_rw_init_one_mr() returns
-EAGAIN. Upper layer protocols treat this as a fatal DMA mapping
failure and tear down the connection.

The max_sgl_rd check is a performance optimization, not a
correctness requirement: the device can handle large SGE counts
via direct posting, just less efficiently than with MR
registration. When the MR pool cannot satisfy a request, falling
back to the direct SGE (map_wrs) path avoids the connection
reset while preserving the MR optimization for the common case
where pool resources are available.

Add a fallback in rdma_rw_ctx_init() so that -EAGAIN from
rdma_rw_init_mr_wrs() triggers direct SGE posting instead of
propagating the error. iWARP devices, which mandate MR
registration for RDMA READs, and force_mr debug mode continue
to treat -EAGAIN as terminal.

Fixes: 00bd1439f464 ("RDMA/rw: Support threshold for registration vs scattering to local pages")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://patch.msgid.link/20260313194201.5818-2-cel@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>