RDMA/core: use IOVA-based DMA mapping for bvec RDMA operations

The bvec RDMA API maps each bvec individually via dma_map_phys(),
requiring an IOTLB sync for each mapping. For large I/O operations
with many bvecs, this overhead becomes significant.

The two-step IOVA API (dma_iova_try_alloc / dma_iova_link /
dma_iova_sync) allocates a contiguous IOVA range upfront, links
all physical pages without IOTLB syncs, then performs a single
sync at the end. This reduces IOTLB flushes from O(n) to O(1).

It also requires only a single output dma_addr_t compared to extra
per-input element storage in struct scatterlist.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Link: https://patch.msgid.link/20260128005400.25147-3-cel@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>

Author: chucklever

drivers/infiniband/core/rw.c

@@ -14,6 +14,7 @@ enum {
 	RDMA_RW_MULTI_WR,
 	RDMA_RW_MR,
 	RDMA_RW_SIG_MR,
+	RDMA_RW_IOVA,
 };
 
 static bool rdma_rw_force_mr;
@@ -383,6 +384,87 @@ static int rdma_rw_init_map_wrs_bvec(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 	return -ENOMEM;
 }
 
+/*
+ * Try to use the two-step IOVA API to map bvecs into a contiguous DMA range.
+ * This reduces IOTLB sync overhead by doing one sync at the end instead of
+ * one per bvec, and produces a contiguous DMA address range that can be
+ * described by a single SGE.
+ *
+ * Returns the number of WQEs (always 1) on success, -EOPNOTSUPP if IOVA
+ * mapping is not available, or another negative error code on failure.
+ */
+static int rdma_rw_init_iova_wrs_bvec(struct rdma_rw_ctx *ctx,
+		struct ib_qp *qp, const struct bio_vec *bvec,
+		struct bvec_iter *iter, u64 remote_addr, u32 rkey,
+		enum dma_data_direction dir)
+{
+	struct ib_device *dev = qp->pd->device;
+	struct device *dma_dev = dev->dma_device;
+	size_t total_len = iter->bi_size;
+	struct bio_vec first_bv;
+	size_t mapped_len = 0;
+	int ret;
+
+	/* Virtual DMA devices cannot support IOVA allocators */
+	if (ib_uses_virt_dma(dev))
+		return -EOPNOTSUPP;
+
+	/* Try to allocate contiguous IOVA space */
+	first_bv = mp_bvec_iter_bvec(bvec, *iter);
+	if (!dma_iova_try_alloc(dma_dev, &ctx->iova.state,
+				bvec_phys(&first_bv), total_len))
+		return -EOPNOTSUPP;
+
+	/* Link all bvecs into the IOVA space */
+	while (iter->bi_size) {
+		struct bio_vec bv = mp_bvec_iter_bvec(bvec, *iter);
+
+		ret = dma_iova_link(dma_dev, &ctx->iova.state, bvec_phys(&bv),
+				    mapped_len, bv.bv_len, dir, 0);
+		if (ret)
+			goto out_destroy;
+
+		mapped_len += bv.bv_len;
+		bvec_iter_advance(bvec, iter, bv.bv_len);
+	}
+
+	/* Sync the IOTLB once for all linked pages */
+	ret = dma_iova_sync(dma_dev, &ctx->iova.state, 0, mapped_len);
+	if (ret)
+		goto out_destroy;
+
+	ctx->iova.mapped_len = mapped_len;
+
+	/* Single SGE covers the entire contiguous IOVA range */
+	ctx->iova.sge.addr = ctx->iova.state.addr;
+	ctx->iova.sge.length = mapped_len;
+	ctx->iova.sge.lkey = qp->pd->local_dma_lkey;
+
+	/* Single WR for the whole transfer */
+	memset(&ctx->iova.wr, 0, sizeof(ctx->iova.wr));
+	if (dir == DMA_TO_DEVICE)
+		ctx->iova.wr.wr.opcode = IB_WR_RDMA_WRITE;
+	else
+		ctx->iova.wr.wr.opcode = IB_WR_RDMA_READ;
+	ctx->iova.wr.wr.num_sge = 1;
+	ctx->iova.wr.wr.sg_list = &ctx->iova.sge;
+	ctx->iova.wr.remote_addr = remote_addr;
+	ctx->iova.wr.rkey = rkey;
+
+	ctx->type = RDMA_RW_IOVA;
+	ctx->nr_ops = 1;
+	return 1;
+
+out_destroy:
+	/*
+	 * dma_iova_destroy() expects the actual mapped length, not the
+	 * total allocation size. It unlinks only the successfully linked
+	 * range and frees the entire IOVA allocation.
+	 */
+	dma_iova_destroy(dma_dev, &ctx->iova.state, mapped_len, dir, 0);
+	return ret;
+}
+
 /**
  * rdma_rw_ctx_init - initialize a RDMA READ/WRITE context
  * @ctx:	context to initialize
@@ -485,6 +567,8 @@ int rdma_rw_ctx_init_bvec(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 		struct bvec_iter iter, u64 remote_addr, u32 rkey,
 		enum dma_data_direction dir)
 {
+	int ret;
+
 	if (nr_bvec == 0 || iter.bi_size == 0)
 		return -EINVAL;
 
@@ -495,6 +579,16 @@ int rdma_rw_ctx_init_bvec(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 	if (nr_bvec == 1)
 		return rdma_rw_init_single_wr_bvec(ctx, qp, bvecs, &iter,
 				remote_addr, rkey, dir);
+
+	/*
+	 * Try IOVA-based mapping first for multi-bvec transfers.
+	 * This reduces IOTLB sync overhead by batching all mappings.
+	 */
+	ret = rdma_rw_init_iova_wrs_bvec(ctx, qp, bvecs, &iter, remote_addr,
+			rkey, dir);
+	if (ret != -EOPNOTSUPP)
+		return ret;
+
 	return rdma_rw_init_map_wrs_bvec(ctx, qp, bvecs, nr_bvec, &iter,
 			remote_addr, rkey, dir);
 }
@@ -671,6 +765,10 @@ struct ib_send_wr *rdma_rw_ctx_wrs(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 			first_wr = &ctx->reg[0].reg_wr.wr;
 		last_wr = &ctx->reg[ctx->nr_ops - 1].wr.wr;
 		break;
+	case RDMA_RW_IOVA:
+		first_wr = &ctx->iova.wr.wr;
+		last_wr = &ctx->iova.wr.wr;
+		break;
 	case RDMA_RW_MULTI_WR:
 		first_wr = &ctx->map.wrs[0].wr;
 		last_wr = &ctx->map.wrs[ctx->nr_ops - 1].wr;
@@ -745,6 +843,10 @@ void rdma_rw_ctx_destroy(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 		break;
 	case RDMA_RW_SINGLE_WR:
 		break;
+	case RDMA_RW_IOVA:
+		/* IOVA contexts must use rdma_rw_ctx_destroy_bvec() */
+		WARN_ON_ONCE(1);
+		return;
 	default:
 		BUG();
 		break;
@@ -778,6 +880,10 @@ void rdma_rw_ctx_destroy_bvec(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
 	u32 i;
 
 	switch (ctx->type) {
+	case RDMA_RW_IOVA:
+		dma_iova_destroy(dev->dma_device, &ctx->iova.state,
+				 ctx->iova.mapped_len, dir, 0);
+		break;
 	case RDMA_RW_MULTI_WR:
 		for (i = 0; i < nr_bvec; i++)
 			ib_dma_unmap_bvec(dev, ctx->map.sges[i].addr,

include/rdma/rw.h

@@ -32,6 +32,14 @@ struct rdma_rw_ctx {
 			struct ib_rdma_wr	*wrs;
 		} map;
 
+		/* for IOVA-based mapping of bvecs into contiguous DMA range: */
+		struct {
+			struct dma_iova_state	state;
+			struct ib_sge		sge;
+			struct ib_rdma_wr	wr;
+			size_t			mapped_len;
+		} iova;
+
 		/* for registering multiple WRs: */
 		struct rdma_rw_reg_ctx {
 			struct ib_sge		sge;