[PATCH for-next v2 0/4] RDMA/erdma: Support non-contiguous kernel queue buffers

[PATCH for-next v2 0/4] RDMA/erdma: Support non-contiguous kernel queue buffers

[Cheng Xu]

Hi,

This small patchset introduces support for non-contiguous kernel queue
buffers. Kernel QP and CQ buffers currently use a single coherent DMA
allocation, which can fail for large queues when memory is fragmented.
To address this, we allocate page-sized coherent buffers and describe
them with the existing MTT.

- #1 adds support for non-contiguous kernel QP buffers.
- #2 adds support for non-contiguous kernel CQ buffers.
- #3 unifies userspace and kernel space queue buffer management.
- #4 moves kernel QP helpers after the shared memory helpers.

Changes from v1:
- Fix formatting issues.

Thanks,
Cheng Xu

Cheng Xu (4):
  RDMA/erdma: Support non-contiguous kernel QP buffers
  RDMA/erdma: Support non-contiguous kernel CQ buffers
  RDMA/erdma: Unify userspace and kernel queue buffer management
  RDMA/erdma: Move kernel QP helpers after memory helpers

 drivers/infiniband/hw/erdma/erdma_cq.c    |  18 +-
 drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +-
 drivers/infiniband/hw/erdma/erdma_verbs.c | 576 +++++++++++++---------
 drivers/infiniband/hw/erdma/erdma_verbs.h |  64 ++-
 4 files changed, 416 insertions(+), 280 deletions(-)

-- 
2.31.1

[PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Cheng Xu]

A single coherent allocation for kernel QP queues can fail for large
queues when memory is fragmented.

Allocate page-sized coherent buffers and describe them with the existing
MTT. Keep the userspace QP path unchanged.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
 drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
 drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
 drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
 4 files changed, 179 insertions(+), 102 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_cq.c b/drivers/infiniband/hw/erdma/erdma_cq.c
index 1f456327e63c..4cc0254b0ab5 100644
--- a/drivers/infiniband/hw/erdma/erdma_cq.c
+++ b/drivers/infiniband/hw/erdma/erdma_cq.c
@@ -161,8 +161,8 @@ static int erdma_poll_one_cqe(struct erdma_cq *cq, struct ib_wc *wc)
 	if (qtype == ERDMA_CQE_QTYPE_SQ) {
 		id_table = kern_qp->swr_tbl;
 		depth = qp->attrs.sq_size;
-		wqe_hdr = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
-					  qp->attrs.sq_size, SQEBB_SHIFT);
+		wqe_hdr = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, wqe_idx,
+					       qp->attrs.sq_size, SQEBB_SHIFT);
 		kern_qp->sq_ci =
 			FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *wqe_hdr) +
 			wqe_idx + 1;
diff --git a/drivers/infiniband/hw/erdma/erdma_qp.c b/drivers/infiniband/hw/erdma/erdma_qp.c
index e002343832f7..fc335b674746 100644
--- a/drivers/infiniband/hw/erdma/erdma_qp.c
+++ b/drivers/infiniband/hw/erdma/erdma_qp.c
@@ -259,8 +259,8 @@ static void erdma_reset_qp(struct erdma_qp *qp)
 	qp->kern_qp.rq_ci = 0;
 	memset(qp->kern_qp.swr_tbl, 0, qp->attrs.sq_size * sizeof(u64));
 	memset(qp->kern_qp.rwr_tbl, 0, qp->attrs.rq_size * sizeof(u64));
-	memset(qp->kern_qp.sq_buf, 0, qp->attrs.sq_size << SQEBB_SHIFT);
-	memset(qp->kern_qp.rq_buf, 0, qp->attrs.rq_size << RQE_SHIFT);
+	erdma_kmem_clear(&qp->kern_qp.sq_mem);
+	erdma_kmem_clear(&qp->kern_qp.rq_mem);
 	erdma_remove_cqes_of_qp(&qp->scq->ibcq, QP_ID(qp));
 	if (qp->rcq != qp->scq)
 		erdma_remove_cqes_of_qp(&qp->rcq->ibcq, QP_ID(qp));
@@ -332,8 +332,8 @@ static int fill_inline_data(struct erdma_qp *qp,
 
 	wqe_idx += (sgl_offset >> SQEBB_SHIFT);
 	sgl_offset &= (SQEBB_SIZE - 1);
-	data = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx, qp->attrs.sq_size,
-			       SQEBB_SHIFT);
+	data = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, wqe_idx,
+				    qp->attrs.sq_size, SQEBB_SHIFT);
 
 	while (i < send_wr->num_sge) {
 		bytes += send_wr->sg_list[i].length;
@@ -356,8 +356,9 @@ static int fill_inline_data(struct erdma_qp *qp,
 			wqe_idx += (sgl_offset >> SQEBB_SHIFT);
 			sgl_offset &= (SQEBB_SIZE - 1);
 
-			data = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
-					       qp->attrs.sq_size, SQEBB_SHIFT);
+			data = erdma_kmem_get_entry(&qp->kern_qp.sq_mem,
+						    wqe_idx, qp->attrs.sq_size,
+						    SQEBB_SHIFT);
 			if (!remain_size)
 				break;
 		}
@@ -385,8 +386,8 @@ static int fill_sgl(struct erdma_qp *qp, const struct ib_send_wr *send_wr,
 	while (i < send_wr->num_sge) {
 		wqe_idx += (sgl_offset >> SQEBB_SHIFT);
 		sgl_offset &= (SQEBB_SIZE - 1);
-		sgl = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
-				      qp->attrs.sq_size, SQEBB_SHIFT);
+		sgl = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, wqe_idx,
+					   qp->attrs.sq_size, SQEBB_SHIFT);
 
 		bytes += send_wr->sg_list[i].length;
 		memcpy(sgl + sgl_offset, &send_wr->sg_list[i],
@@ -463,8 +464,8 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 	    send_wr->opcode != IB_WR_SEND_WITH_IMM)
 		return -EINVAL;
 
-	entry = get_queue_entry(qp->kern_qp.sq_buf, idx, qp->attrs.sq_size,
-				SQEBB_SHIFT);
+	entry = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, idx,
+				     qp->attrs.sq_size, SQEBB_SHIFT);
 
 	/* Clear the SQE header section. */
 	*entry = 0;
@@ -524,8 +525,8 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 		read_sqe->sink_to_h =
 			cpu_to_le32(upper_32_bits(send_wr->sg_list[0].addr));
 
-		sge = get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
-				      qp->attrs.sq_size, SQEBB_SHIFT);
+		sge = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, idx + 1,
+					   qp->attrs.sq_size, SQEBB_SHIFT);
 		sge->addr = cpu_to_le64(rdma_wr->remote_addr);
 		sge->key = cpu_to_le32(rdma_wr->rkey);
 		sge->length = cpu_to_le32(send_wr->sg_list[0].length);
@@ -569,8 +570,9 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 		if (mr->mem.mtt_nents <= ERDMA_MAX_INLINE_MTT_ENTRIES) {
 			attrs |= FIELD_PREP(ERDMA_SQE_MR_MTT_TYPE_MASK, 0);
 			/* Copy SGLs to SQE content to accelerate */
-			memcpy(get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
-					       qp->attrs.sq_size, SQEBB_SHIFT),
+			memcpy(erdma_kmem_get_entry(&qp->kern_qp.sq_mem,
+						    idx + 1, qp->attrs.sq_size,
+						    SQEBB_SHIFT),
 			       mr->mem.mtt->buf, MTT_SIZE(mr->mem.mtt_nents));
 			wqe_size = sizeof(struct erdma_reg_mr_sqe) +
 				   MTT_SIZE(mr->mem.mtt_nents);
@@ -605,8 +607,8 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 				cpu_to_le64(atomic_wr(send_wr)->compare_add);
 		}
 
-		sge = get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
-				      qp->attrs.sq_size, SQEBB_SHIFT);
+		sge = erdma_kmem_get_entry(&qp->kern_qp.sq_mem, idx + 1,
+					   qp->attrs.sq_size, SQEBB_SHIFT);
 		sge->addr = cpu_to_le64(atomic_wr(send_wr)->remote_addr);
 		sge->key = cpu_to_le32(atomic_wr(send_wr)->rkey);
 		sge++;
@@ -702,8 +704,8 @@ static int erdma_post_recv_one(struct erdma_qp *qp,
 			       const struct ib_recv_wr *recv_wr)
 {
 	struct erdma_rqe *rqe =
-		get_queue_entry(qp->kern_qp.rq_buf, qp->kern_qp.rq_pi,
-				qp->attrs.rq_size, RQE_SHIFT);
+		erdma_kmem_get_entry(&qp->kern_qp.rq_mem, qp->kern_qp.rq_pi,
+				     qp->attrs.rq_size, RQE_SHIFT);
 
 	rqe->qe_idx = cpu_to_le16(qp->kern_qp.rq_pi + 1);
 	rqe->qpn = cpu_to_le32(QP_ID(qp));
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index 65b1af1e6623..a2de700bce90 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -19,6 +19,10 @@
 #include "erdma_cm.h"
 #include "erdma_verbs.h"
 
+static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
+			    size_t size);
+static void put_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem);
+
 static void assemble_qbuf_mtt_for_cmd(struct erdma_mem *mem, u32 *cfg,
 				      u64 *addr0, u64 *addr1)
 {
@@ -41,7 +45,7 @@ static int create_qp_cmd(struct erdma_ucontext *uctx, struct erdma_qp *qp)
 	struct erdma_dev *dev = to_edev(qp->ibqp.device);
 	struct erdma_pd *pd = to_epd(qp->ibqp.pd);
 	struct erdma_cmdq_create_qp_req req;
-	struct erdma_uqp *user_qp;
+	struct erdma_mem *sq_mem, *rq_mem;
 	u64 resp0, resp1;
 	int err;
 
@@ -63,67 +67,46 @@ static int create_qp_cmd(struct erdma_ucontext *uctx, struct erdma_qp *qp)
 				      ERDMA_QPT_UD);
 
 	if (rdma_is_kernel_res(&qp->ibqp.res)) {
-		u32 pgsz_range = ilog2(SZ_1M) - ERDMA_HW_PAGE_SHIFT;
-
-		req.sq_cqn_mtt_cfg =
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
-				   pgsz_range) |
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->scq->cqn);
-		req.rq_cqn_mtt_cfg =
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
-				   pgsz_range) |
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->rcq->cqn);
-
-		req.sq_mtt_cfg =
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_OFFSET_MASK, 0) |
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK, 1) |
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_LEVEL_MASK,
-				   ERDMA_MR_MTT_0LEVEL);
-		req.rq_mtt_cfg = req.sq_mtt_cfg;
-
-		req.rq_buf_addr = qp->kern_qp.rq_buf_dma_addr;
-		req.sq_buf_addr = qp->kern_qp.sq_buf_dma_addr;
+		sq_mem = &qp->kern_qp.sq_mem;
+		rq_mem = &qp->kern_qp.rq_mem;
 		req.sq_dbrec_dma = qp->kern_qp.sq_dbrec_dma;
 		req.rq_dbrec_dma = qp->kern_qp.rq_dbrec_dma;
 	} else {
-		user_qp = &qp->user_qp;
-		req.sq_cqn_mtt_cfg = FIELD_PREP(
-			ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
-			ilog2(user_qp->sq_mem.page_size) - ERDMA_HW_PAGE_SHIFT);
-		req.sq_cqn_mtt_cfg |=
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->scq->cqn);
-
-		req.rq_cqn_mtt_cfg = FIELD_PREP(
-			ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
-			ilog2(user_qp->rq_mem.page_size) - ERDMA_HW_PAGE_SHIFT);
-		req.rq_cqn_mtt_cfg |=
-			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->rcq->cqn);
-
-		req.sq_mtt_cfg = user_qp->sq_mem.page_offset;
-		req.sq_mtt_cfg |= FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK,
-					     user_qp->sq_mem.mtt_nents);
-
-		req.rq_mtt_cfg = user_qp->rq_mem.page_offset;
-		req.rq_mtt_cfg |= FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK,
-					     user_qp->rq_mem.mtt_nents);
-
-		assemble_qbuf_mtt_for_cmd(&user_qp->sq_mem, &req.sq_mtt_cfg,
-					  &req.sq_buf_addr, req.sq_mtt_entry);
-		assemble_qbuf_mtt_for_cmd(&user_qp->rq_mem, &req.rq_mtt_cfg,
-					  &req.rq_buf_addr, req.rq_mtt_entry);
-
-		req.sq_dbrec_dma = user_qp->sq_dbrec_dma;
-		req.rq_dbrec_dma = user_qp->rq_dbrec_dma;
+		sq_mem = &qp->user_qp.sq_mem;
+		rq_mem = &qp->user_qp.rq_mem;
+		req.sq_dbrec_dma = qp->user_qp.sq_dbrec_dma;
+		req.rq_dbrec_dma = qp->user_qp.rq_dbrec_dma;
+	}
 
-		if (uctx->ext_db.enable) {
-			req.sq_cqn_mtt_cfg |=
-				FIELD_PREP(ERDMA_CMD_CREATE_QP_DB_CFG_MASK, 1);
-			req.db_cfg =
-				FIELD_PREP(ERDMA_CMD_CREATE_QP_SQDB_CFG_MASK,
-					   uctx->ext_db.sdb_off) |
-				FIELD_PREP(ERDMA_CMD_CREATE_QP_RQDB_CFG_MASK,
-					   uctx->ext_db.rdb_off);
-		}
+	req.sq_cqn_mtt_cfg =
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
+			   ilog2(sq_mem->page_size) - ERDMA_HW_PAGE_SHIFT);
+	req.sq_cqn_mtt_cfg |=
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->scq->cqn);
+	req.rq_cqn_mtt_cfg =
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
+			   ilog2(rq_mem->page_size) - ERDMA_HW_PAGE_SHIFT);
+	req.rq_cqn_mtt_cfg |=
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->rcq->cqn);
+
+	req.sq_mtt_cfg =
+		sq_mem->page_offset |
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK, sq_mem->mtt_nents);
+	req.rq_mtt_cfg =
+		rq_mem->page_offset |
+		FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK, rq_mem->mtt_nents);
+
+	assemble_qbuf_mtt_for_cmd(sq_mem, &req.sq_mtt_cfg, &req.sq_buf_addr,
+				  req.sq_mtt_entry);
+	assemble_qbuf_mtt_for_cmd(rq_mem, &req.rq_mtt_cfg, &req.rq_buf_addr,
+				  req.rq_mtt_entry);
+	if (uctx && uctx->ext_db.enable) {
+		req.sq_cqn_mtt_cfg |=
+			FIELD_PREP(ERDMA_CMD_CREATE_QP_DB_CFG_MASK, 1);
+		req.db_cfg = FIELD_PREP(ERDMA_CMD_CREATE_QP_SQDB_CFG_MASK,
+					uctx->ext_db.sdb_off) |
+			     FIELD_PREP(ERDMA_CMD_CREATE_QP_RQDB_CFG_MASK,
+					uctx->ext_db.rdb_off);
 	}
 
 	err = erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), &resp0, &resp1,
@@ -512,21 +495,13 @@ static void free_kernel_qp(struct erdma_qp *qp)
 	vfree(qp->kern_qp.swr_tbl);
 	vfree(qp->kern_qp.rwr_tbl);
 
-	if (qp->kern_qp.sq_buf)
-		dma_free_coherent(&dev->pdev->dev,
-				  qp->attrs.sq_size << SQEBB_SHIFT,
-				  qp->kern_qp.sq_buf,
-				  qp->kern_qp.sq_buf_dma_addr);
+	put_mtt_entries(dev, &qp->kern_qp.sq_mem);
 
 	if (qp->kern_qp.sq_dbrec)
 		dma_pool_free(dev->db_pool, qp->kern_qp.sq_dbrec,
 			      qp->kern_qp.sq_dbrec_dma);
 
-	if (qp->kern_qp.rq_buf)
-		dma_free_coherent(&dev->pdev->dev,
-				  qp->attrs.rq_size << RQE_SHIFT,
-				  qp->kern_qp.rq_buf,
-				  qp->kern_qp.rq_buf_dma_addr);
+	put_mtt_entries(dev, &qp->kern_qp.rq_mem);
 
 	if (qp->kern_qp.rq_dbrec)
 		dma_pool_free(dev->db_pool, qp->kern_qp.rq_dbrec,
@@ -556,9 +531,7 @@ static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
 		goto err_out;
 
 	size = qp->attrs.sq_size << SQEBB_SHIFT;
-	kqp->sq_buf = dma_alloc_coherent(&dev->pdev->dev, size,
-					 &kqp->sq_buf_dma_addr, GFP_KERNEL);
-	if (!kqp->sq_buf)
+	if (erdma_alloc_kmem(dev, &kqp->sq_mem, size))
 		goto err_out;
 
 	kqp->sq_dbrec =
@@ -567,9 +540,7 @@ static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
 		goto err_out;
 
 	size = qp->attrs.rq_size << RQE_SHIFT;
-	kqp->rq_buf = dma_alloc_coherent(&dev->pdev->dev, size,
-					 &kqp->rq_buf_dma_addr, GFP_KERNEL);
-	if (!kqp->rq_buf)
+	if (erdma_alloc_kmem(dev, &kqp->rq_mem, size))
 		goto err_out;
 
 	kqp->rq_dbrec =
@@ -826,12 +797,80 @@ static void erdma_destroy_mtt(struct erdma_dev *dev, struct erdma_mtt *mtt)
 	}
 }
 
+static void erdma_free_kmem(struct erdma_dev *dev, struct erdma_mem *mem)
+{
+	struct erdma_buf_list *buf_list;
+	u32 i;
+
+	if (!mem->kmem)
+		return;
+
+	buf_list = mem->kmem->buf_list;
+	if (buf_list) {
+		for (i = 0; i < mem->page_cnt; i++)
+			if (buf_list[i].buf)
+				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+						  buf_list[i].buf,
+						  buf_list[i].dma_addr);
+		kfree(buf_list);
+	}
+
+	kfree(mem->kmem);
+	mem->kmem = NULL;
+}
+
+static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
+			    size_t size)
+{
+	struct erdma_buf_list *buf_list;
+	u32 i;
+
+	mem->type = ERDMA_KMEM;
+	mem->page_size = PAGE_SIZE;
+	mem->page_offset = 0;
+	mem->page_cnt = DIV_ROUND_UP(size, PAGE_SIZE);
+	mem->mtt_nents = mem->page_cnt;
+	mem->len = size;
+
+	mem->kmem = kzalloc_obj(*mem->kmem);
+	if (!mem->kmem)
+		goto err_free_mem;
+
+	buf_list = kcalloc(mem->page_cnt, sizeof(*buf_list), GFP_KERNEL);
+	if (!buf_list)
+		goto err_free_mem;
+	mem->kmem->buf_list = buf_list;
+
+	for (i = 0; i < mem->page_cnt; i++) {
+		buf_list[i].buf = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+						     &buf_list[i].dma_addr,
+						     GFP_KERNEL);
+		if (!buf_list[i].buf)
+			goto err_free_mem;
+	}
+
+	mem->mtt = erdma_create_mtt(dev, MTT_SIZE(mem->page_cnt), true);
+	if (IS_ERR(mem->mtt)) {
+		mem->mtt = NULL;
+		goto err_free_mem;
+	}
+	for (i = 0; i < mem->page_cnt; i++)
+		mem->mtt->buf[i] = buf_list[i].dma_addr;
+
+	return 0;
+
+err_free_mem:
+	erdma_free_kmem(dev, mem);
+	return -ENOMEM;
+}
+
 static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
 			   u64 start, u64 len, int access, u64 virt,
 			   unsigned long req_page_size, bool force_continuous)
 {
 	int ret = 0;
 
+	mem->type = ERDMA_UMEM;
 	mem->umem = ib_umem_get_va(&dev->ibdev, start, len, access);
 	if (IS_ERR(mem->umem)) {
 		ret = PTR_ERR(mem->umem);
@@ -867,10 +906,14 @@ static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
 
 static void put_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem)
 {
-	if (mem->mtt)
+	if (mem->mtt) {
 		erdma_destroy_mtt(dev, mem->mtt);
+		mem->mtt = NULL;
+	}
 
-	if (mem->umem) {
+	if (mem->type == ERDMA_KMEM) {
+		erdma_free_kmem(dev, mem);
+	} else if (mem->umem) {
 		ib_umem_release(mem->umem);
 		mem->umem = NULL;
 	}
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.h b/drivers/infiniband/hw/erdma/erdma_verbs.h
index c73cecf92f61..eb7318cf7a52 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.h
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.h
@@ -111,8 +111,26 @@ struct erdma_mtt {
 	struct erdma_mtt *low_level;
 };
 
+enum erdma_mem_type {
+	ERDMA_UMEM = 0,
+	ERDMA_KMEM = 1,
+};
+
+struct erdma_buf_list {
+	void *buf;
+	dma_addr_t dma_addr;
+};
+
+struct erdma_kmem {
+	struct erdma_buf_list *buf_list;
+};
+
 struct erdma_mem {
-	struct ib_umem *umem;
+	enum erdma_mem_type type;
+	union {
+		struct ib_umem *umem;
+		struct erdma_kmem *kmem;
+	};
 	struct erdma_mtt *mtt;
 
 	u32 page_size;
@@ -124,6 +142,23 @@ struct erdma_mem {
 	u64 len;
 };
 
+static inline void *erdma_kmem_get_entry(struct erdma_mem *mem, u32 idx,
+					 u32 depth, u32 shift)
+{
+	u32 offset = (idx & (depth - 1)) << shift;
+
+	return (u8 *)mem->kmem->buf_list[offset >> PAGE_SHIFT].buf +
+	       offset_in_page(offset);
+}
+
+static inline void erdma_kmem_clear(struct erdma_mem *mem)
+{
+	u32 i;
+
+	for (i = 0; i < mem->page_cnt; i++)
+		memset(mem->kmem->buf_list[i].buf, 0, PAGE_SIZE);
+}
+
 struct erdma_mr {
 	struct ib_mr ibmr;
 	struct erdma_mem mem;
@@ -183,11 +218,8 @@ struct erdma_kqp {
 	void __iomem *hw_sq_db;
 	void __iomem *hw_rq_db;
 
-	void *sq_buf;
-	dma_addr_t sq_buf_dma_addr;
-
-	void *rq_buf;
-	dma_addr_t rq_buf_dma_addr;
+	struct erdma_mem sq_mem;
+	struct erdma_mem rq_mem;
 
 	void *sq_dbrec;
 	void *rq_dbrec;
-- 
2.31.1

[PATCH for-next v2 2/4] RDMA/erdma: Support non-contiguous kernel CQ buffers

[Cheng Xu]

A single coherent allocation for a kernel CQ can fail when memory is
fragmented.

Use page-sized coherent buffers and describe them with the existing MTT.
Keep the userspace CQ path and doorbell allocation unchanged.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma_cq.c    | 14 ++---
 drivers/infiniband/hw/erdma/erdma_verbs.c | 76 +++++++++--------------
 drivers/infiniband/hw/erdma/erdma_verbs.h |  3 +-
 3 files changed, 36 insertions(+), 57 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_cq.c b/drivers/infiniband/hw/erdma/erdma_cq.c
index 4cc0254b0ab5..daec74f0aa6d 100644
--- a/drivers/infiniband/hw/erdma/erdma_cq.c
+++ b/drivers/infiniband/hw/erdma/erdma_cq.c
@@ -8,8 +8,8 @@
 
 static void *get_next_valid_cqe(struct erdma_cq *cq)
 {
-	__be32 *cqe = get_queue_entry(cq->kern_cq.qbuf, cq->kern_cq.ci,
-				      cq->depth, CQE_SHIFT);
+	__be32 *cqe = erdma_kmem_get_entry(
+		&cq->kern_cq.qbuf_mem, cq->kern_cq.ci, cq->depth, CQE_SHIFT);
 	u32 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
 			      be32_to_cpu(READ_ONCE(*cqe)));
 
@@ -242,15 +242,15 @@ void erdma_remove_cqes_of_qp(struct ib_cq *ibcq, u32 qpn)
 
 	while (ncqe > 0) {
 		cur_cq_ci = prev_cq_ci + ncqe - 1;
-		cqe = get_queue_entry(cq->kern_cq.qbuf, cur_cq_ci, cq->depth,
-				      CQE_SHIFT);
+		cqe = erdma_kmem_get_entry(&cq->kern_cq.qbuf_mem, cur_cq_ci,
+					   cq->depth, CQE_SHIFT);
 
 		if (be32_to_cpu(cqe->qpn) == qpn) {
 			++nqp_cqe;
 		} else if (nqp_cqe) {
-			dst_cqe = get_queue_entry(cq->kern_cq.qbuf,
-						  cur_cq_ci + nqp_cqe,
-						  cq->depth, CQE_SHIFT);
+			dst_cqe = erdma_kmem_get_entry(&cq->kern_cq.qbuf_mem,
+						       cur_cq_ci + nqp_cqe,
+						       cq->depth, CQE_SHIFT);
 			owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
 					  be32_to_cpu(dst_cqe->hdr));
 			cqe->hdr = cpu_to_be32(
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index a2de700bce90..b0a47c82cbdf 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -177,7 +177,6 @@ static int create_cq_cmd(struct erdma_ucontext *uctx, struct erdma_cq *cq)
 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
 	struct erdma_cmdq_create_cq_req req;
 	struct erdma_mem *mem;
-	u32 page_size;
 
 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
 				CMDQ_OPCODE_CREATE_CQ);
@@ -187,48 +186,34 @@ static int create_cq_cmd(struct erdma_ucontext *uctx, struct erdma_cq *cq)
 	req.cfg1 = FIELD_PREP(ERDMA_CMD_CREATE_CQ_EQN_MASK, cq->assoc_eqn);
 
 	if (rdma_is_kernel_res(&cq->ibcq.res)) {
-		page_size = SZ_32M;
-		req.cfg0 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK,
-				       ilog2(page_size) - ERDMA_HW_PAGE_SHIFT);
-		req.qbuf_addr_l = lower_32_bits(cq->kern_cq.qbuf_dma_addr);
-		req.qbuf_addr_h = upper_32_bits(cq->kern_cq.qbuf_dma_addr);
-
-		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK, 1) |
-			    FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_LEVEL_MASK,
-				       ERDMA_MR_MTT_0LEVEL);
-
-		req.first_page_offset = 0;
+		mem = &cq->kern_cq.qbuf_mem;
 		req.cq_dbrec_dma = cq->kern_cq.dbrec_dma;
 	} else {
 		mem = &cq->user_cq.qbuf_mem;
-		req.cfg0 |=
-			FIELD_PREP(ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK,
-				   ilog2(mem->page_size) - ERDMA_HW_PAGE_SHIFT);
-		if (mem->mtt_nents == 1) {
-			req.qbuf_addr_l = lower_32_bits(mem->mtt->buf[0]);
-			req.qbuf_addr_h = upper_32_bits(mem->mtt->buf[0]);
-			req.cfg1 |=
-				FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_LEVEL_MASK,
-					   ERDMA_MR_MTT_0LEVEL);
-		} else {
-			req.qbuf_addr_l = lower_32_bits(mem->mtt->buf_dma);
-			req.qbuf_addr_h = upper_32_bits(mem->mtt->buf_dma);
-			req.cfg1 |=
-				FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_LEVEL_MASK,
-					   ERDMA_MR_MTT_1LEVEL);
-		}
-		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK,
-				       mem->mtt_nents);
-
-		req.first_page_offset = mem->page_offset;
 		req.cq_dbrec_dma = cq->user_cq.dbrec_dma;
+	}
 
-		if (uctx->ext_db.enable) {
-			req.cfg1 |= FIELD_PREP(
-				ERDMA_CMD_CREATE_CQ_MTT_DB_CFG_MASK, 1);
-			req.cfg2 = FIELD_PREP(ERDMA_CMD_CREATE_CQ_DB_CFG_MASK,
-					      uctx->ext_db.cdb_off);
-		}
+	req.cfg0 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK,
+			       ilog2(mem->page_size) - ERDMA_HW_PAGE_SHIFT);
+	if (mem->mtt_nents == 1) {
+		req.qbuf_addr_l = lower_32_bits(mem->mtt->buf[0]);
+		req.qbuf_addr_h = upper_32_bits(mem->mtt->buf[0]);
+		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_LEVEL_MASK,
+				       ERDMA_MR_MTT_0LEVEL);
+	} else {
+		req.qbuf_addr_l = lower_32_bits(mem->mtt->buf_dma);
+		req.qbuf_addr_h = upper_32_bits(mem->mtt->buf_dma);
+		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_LEVEL_MASK,
+				       ERDMA_MR_MTT_1LEVEL);
+	}
+	req.cfg1 |=
+		FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK, mem->mtt_nents);
+	req.first_page_offset = mem->page_offset;
+
+	if (uctx && uctx->ext_db.enable) {
+		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_DB_CFG_MASK, 1);
+		req.cfg2 = FIELD_PREP(ERDMA_CMD_CREATE_CQ_DB_CFG_MASK,
+				      uctx->ext_db.cdb_off);
 	}
 
 	return erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL,
@@ -1392,8 +1377,7 @@ int erdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 	wait_for_completion(&cq->free);
 
 	if (rdma_is_kernel_res(&cq->ibcq.res)) {
-		dma_free_coherent(&dev->pdev->dev, cq->depth << CQE_SHIFT,
-				  cq->kern_cq.qbuf, cq->kern_cq.qbuf_dma_addr);
+		put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
 		dma_pool_free(dev->db_pool, cq->kern_cq.dbrec,
 			      cq->kern_cq.dbrec_dma);
 	} else {
@@ -1991,10 +1975,8 @@ static int erdma_init_kernel_cq(struct erdma_cq *cq)
 {
 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
 
-	cq->kern_cq.qbuf =
-		dma_alloc_coherent(&dev->pdev->dev, cq->depth << CQE_SHIFT,
-				   &cq->kern_cq.qbuf_dma_addr, GFP_KERNEL);
-	if (!cq->kern_cq.qbuf)
+	if (erdma_alloc_kmem(dev, &cq->kern_cq.qbuf_mem,
+			     cq->depth << CQE_SHIFT))
 		return -ENOMEM;
 
 	cq->kern_cq.dbrec = dma_pool_zalloc(dev->db_pool, GFP_KERNEL,
@@ -2009,8 +1991,7 @@ static int erdma_init_kernel_cq(struct erdma_cq *cq)
 	return 0;
 
 err_out:
-	dma_free_coherent(&dev->pdev->dev, cq->depth << CQE_SHIFT,
-			  cq->kern_cq.qbuf, cq->kern_cq.qbuf_dma_addr);
+	put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
 
 	return -ENOMEM;
 }
@@ -2077,8 +2058,7 @@ int erdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		erdma_unmap_user_dbrecords(ctx, &cq->user_cq.user_dbr_page);
 		put_mtt_entries(dev, &cq->user_cq.qbuf_mem);
 	} else {
-		dma_free_coherent(&dev->pdev->dev, depth << CQE_SHIFT,
-				  cq->kern_cq.qbuf, cq->kern_cq.qbuf_dma_addr);
+		put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
 		dma_pool_free(dev->db_pool, cq->kern_cq.dbrec,
 			      cq->kern_cq.dbrec_dma);
 	}
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.h b/drivers/infiniband/hw/erdma/erdma_verbs.h
index eb7318cf7a52..5056d60357db 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.h
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.h
@@ -352,8 +352,7 @@ struct erdma_qp {
 };
 
 struct erdma_kcq_info {
-	void *qbuf;
-	dma_addr_t qbuf_dma_addr;
+	struct erdma_mem qbuf_mem;
 	u32 ci;
 	u32 cmdsn;
 	u32 notify_cnt;
-- 
2.31.1

[PATCH for-next v2 3/4] RDMA/erdma: Unify userspace and kernel queue buffer management

[Cheng Xu]

Userspace and kernel space queue buffers use separate helpers despite
sharing MTT metadata and lifetime rules. Manage both through
erdma_mem_init() and erdma_mem_uninit(), while keeping backing allocation
and release type-specific.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma_verbs.c | 281 ++++++++++++++--------
 drivers/infiniband/hw/erdma/erdma_verbs.h |  17 +-
 2 files changed, 191 insertions(+), 107 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index b0a47c82cbdf..aaef2b3621b0 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -19,9 +19,9 @@
 #include "erdma_cm.h"
 #include "erdma_verbs.h"
 
-static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
-			    size_t size);
-static void put_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem);
+static int erdma_mem_init(struct erdma_dev *dev, struct erdma_mem *mem,
+			  struct erdma_mem_init_attr *attr);
+static void erdma_mem_uninit(struct erdma_dev *dev, struct erdma_mem *mem);
 
 static void assemble_qbuf_mtt_for_cmd(struct erdma_mem *mem, u32 *cfg,
 				      u64 *addr0, u64 *addr1)
@@ -480,13 +480,13 @@ static void free_kernel_qp(struct erdma_qp *qp)
 	vfree(qp->kern_qp.swr_tbl);
 	vfree(qp->kern_qp.rwr_tbl);
 
-	put_mtt_entries(dev, &qp->kern_qp.sq_mem);
+	erdma_mem_uninit(dev, &qp->kern_qp.sq_mem);
 
 	if (qp->kern_qp.sq_dbrec)
 		dma_pool_free(dev->db_pool, qp->kern_qp.sq_dbrec,
 			      qp->kern_qp.sq_dbrec_dma);
 
-	put_mtt_entries(dev, &qp->kern_qp.rq_mem);
+	erdma_mem_uninit(dev, &qp->kern_qp.rq_mem);
 
 	if (qp->kern_qp.rq_dbrec)
 		dma_pool_free(dev->db_pool, qp->kern_qp.rq_dbrec,
@@ -496,8 +496,12 @@ static void free_kernel_qp(struct erdma_qp *qp)
 static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
 			  struct ib_qp_init_attr *attrs)
 {
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_KMEM,
+	};
 	struct erdma_kqp *kqp = &qp->kern_qp;
 	int size;
+	int ret;
 
 	if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
 		kqp->sig_all = 1;
@@ -512,45 +516,66 @@ static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
 
 	kqp->swr_tbl = vmalloc_array(qp->attrs.sq_size, sizeof(u64));
 	kqp->rwr_tbl = vmalloc_array(qp->attrs.rq_size, sizeof(u64));
-	if (!kqp->swr_tbl || !kqp->rwr_tbl)
+	if (!kqp->swr_tbl || !kqp->rwr_tbl) {
+		ret = -ENOMEM;
 		goto err_out;
+	}
 
 	size = qp->attrs.sq_size << SQEBB_SHIFT;
-	if (erdma_alloc_kmem(dev, &kqp->sq_mem, size))
+	attr.len = size;
+	ret = erdma_mem_init(dev, &kqp->sq_mem, &attr);
+	if (ret)
 		goto err_out;
 
 	kqp->sq_dbrec =
 		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->sq_dbrec_dma);
-	if (!kqp->sq_dbrec)
+	if (!kqp->sq_dbrec) {
+		ret = -ENOMEM;
 		goto err_out;
+	}
 
 	size = qp->attrs.rq_size << RQE_SHIFT;
-	if (erdma_alloc_kmem(dev, &kqp->rq_mem, size))
+	attr.len = size;
+	ret = erdma_mem_init(dev, &kqp->rq_mem, &attr);
+	if (ret)
 		goto err_out;
 
 	kqp->rq_dbrec =
 		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->rq_dbrec_dma);
-	if (!kqp->rq_dbrec)
+	if (!kqp->rq_dbrec) {
+		ret = -ENOMEM;
 		goto err_out;
+	}
 
 	return 0;
 
 err_out:
 	free_kernel_qp(qp);
-	return -ENOMEM;
+	return ret;
 }
 
-static void erdma_fill_bottom_mtt(struct erdma_dev *dev, struct erdma_mem *mem)
+static void erdma_init_mtt_leaf(struct erdma_mem *mem, struct erdma_mtt *mtt)
 {
-	struct erdma_mtt *mtt = mem->mtt;
 	struct ib_block_iter biter;
 	u32 idx = 0;
 
+	if (mem->type == ERDMA_UMEM) {
+		rdma_umem_for_each_dma_block(mem->umem, &biter, mem->page_size)
+			mtt->buf[idx++] = rdma_block_iter_dma_address(&biter);
+	} else {
+		for (; idx < mem->page_cnt; idx++)
+			mtt->buf[idx] = mem->kmem->buf_list[idx].dma_addr;
+	}
+}
+
+static void erdma_init_bottom_mtt(struct erdma_dev *dev, struct erdma_mem *mem)
+{
+	struct erdma_mtt *mtt = mem->mtt;
+
 	while (mtt->low_level)
 		mtt = mtt->low_level;
 
-	rdma_umem_for_each_dma_block(mem->umem, &biter, mem->page_size)
-		mtt->buf[idx++] = rdma_block_iter_dma_address(&biter);
+	erdma_init_mtt_leaf(mem, mtt);
 }
 
 static struct erdma_mtt *erdma_create_cont_mtt(struct erdma_dev *dev,
@@ -782,7 +807,7 @@ static void erdma_destroy_mtt(struct erdma_dev *dev, struct erdma_mtt *mtt)
 	}
 }
 
-static void erdma_free_kmem(struct erdma_dev *dev, struct erdma_mem *mem)
+static void erdma_kmem_free(struct erdma_dev *dev, struct erdma_mem *mem)
 {
 	struct erdma_buf_list *buf_list;
 	u32 i;
@@ -804,104 +829,126 @@ static void erdma_free_kmem(struct erdma_dev *dev, struct erdma_mem *mem)
 	mem->kmem = NULL;
 }
 
-static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
-			    size_t size)
+static int erdma_kmem_alloc(struct erdma_dev *dev, struct erdma_mem *mem,
+			    u64 size)
 {
 	struct erdma_buf_list *buf_list;
+	u32 page_cnt = DIV_ROUND_UP(size, PAGE_SIZE);
 	u32 i;
 
-	mem->type = ERDMA_KMEM;
-	mem->page_size = PAGE_SIZE;
-	mem->page_offset = 0;
-	mem->page_cnt = DIV_ROUND_UP(size, PAGE_SIZE);
-	mem->mtt_nents = mem->page_cnt;
-	mem->len = size;
-
 	mem->kmem = kzalloc_obj(*mem->kmem);
 	if (!mem->kmem)
-		goto err_free_mem;
+		return -ENOMEM;
 
-	buf_list = kcalloc(mem->page_cnt, sizeof(*buf_list), GFP_KERNEL);
+	buf_list = kcalloc(page_cnt, sizeof(*buf_list), GFP_KERNEL);
 	if (!buf_list)
-		goto err_free_mem;
+		goto err_free_kmem;
 	mem->kmem->buf_list = buf_list;
 
-	for (i = 0; i < mem->page_cnt; i++) {
+	for (i = 0; i < page_cnt; i++) {
 		buf_list[i].buf = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
 						     &buf_list[i].dma_addr,
 						     GFP_KERNEL);
 		if (!buf_list[i].buf)
-			goto err_free_mem;
+			goto err_free_pages;
 	}
 
-	mem->mtt = erdma_create_mtt(dev, MTT_SIZE(mem->page_cnt), true);
-	if (IS_ERR(mem->mtt)) {
-		mem->mtt = NULL;
-		goto err_free_mem;
-	}
-	for (i = 0; i < mem->page_cnt; i++)
-		mem->mtt->buf[i] = buf_list[i].dma_addr;
-
 	return 0;
 
-err_free_mem:
-	erdma_free_kmem(dev, mem);
+err_free_pages:
+	while (i--)
+		dma_free_coherent(&dev->pdev->dev, PAGE_SIZE, buf_list[i].buf,
+				  buf_list[i].dma_addr);
+	kfree(buf_list);
+
+err_free_kmem:
+	kfree(mem->kmem);
+	mem->kmem = NULL;
+
 	return -ENOMEM;
 }
 
-static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
-			   u64 start, u64 len, int access, u64 virt,
-			   unsigned long req_page_size, bool force_continuous)
+static void erdma_mem_free(struct erdma_dev *dev, struct erdma_mem *mem)
 {
-	int ret = 0;
+	switch (mem->type) {
+	case ERDMA_UMEM:
+		if (mem->umem) {
+			ib_umem_release(mem->umem);
+			mem->umem = NULL;
+		}
+		break;
+	case ERDMA_KMEM:
+		erdma_kmem_free(dev, mem);
+		break;
+	default:
+		break;
+	}
+}
 
-	mem->type = ERDMA_UMEM;
-	mem->umem = ib_umem_get_va(&dev->ibdev, start, len, access);
-	if (IS_ERR(mem->umem)) {
-		ret = PTR_ERR(mem->umem);
-		mem->umem = NULL;
-		return ret;
+static int erdma_mem_init(struct erdma_dev *dev, struct erdma_mem *mem,
+			  struct erdma_mem_init_attr *attr)
+{
+	int ret;
+
+	mem->type = attr->type;
+
+	switch (mem->type) {
+	case ERDMA_UMEM:
+		mem->umem = ib_umem_get_va(&dev->ibdev, attr->start, attr->len,
+					   attr->access);
+		if (IS_ERR(mem->umem)) {
+			ret = PTR_ERR(mem->umem);
+			mem->umem = NULL;
+			return ret;
+		}
+
+		mem->page_size = ib_umem_find_best_pgsz(
+			mem->umem, attr->req_page_size, attr->virt);
+		mem->mtt_nents =
+			ib_umem_num_dma_blocks(mem->umem, mem->page_size);
+		break;
+	case ERDMA_KMEM:
+		ret = erdma_kmem_alloc(dev, mem, attr->len);
+		if (ret)
+			return ret;
+
+		mem->page_size = PAGE_SIZE;
+		mem->mtt_nents = DIV_ROUND_UP(attr->len, PAGE_SIZE);
+		break;
+	default:
+		return -EINVAL;
 	}
 
-	mem->va = virt;
-	mem->len = len;
-	mem->page_size = ib_umem_find_best_pgsz(mem->umem, req_page_size, virt);
-	mem->page_offset = start & (mem->page_size - 1);
-	mem->mtt_nents = ib_umem_num_dma_blocks(mem->umem, mem->page_size);
+	mem->va = attr->virt;
+	mem->len = attr->len;
+	mem->page_offset = attr->start & (mem->page_size - 1);
 	mem->page_cnt = mem->mtt_nents;
 	mem->mtt = erdma_create_mtt(dev, MTT_SIZE(mem->page_cnt),
-				    force_continuous);
+				    !(attr->flags & ERDMA_MEM_FLAG_MR_BUF));
 	if (IS_ERR(mem->mtt)) {
 		ret = PTR_ERR(mem->mtt);
-		goto error_ret;
+		mem->mtt = NULL;
+		goto err_free_mem;
 	}
 
-	erdma_fill_bottom_mtt(dev, mem);
+	erdma_init_bottom_mtt(dev, mem);
 
 	return 0;
 
-error_ret:
-	if (mem->umem) {
-		ib_umem_release(mem->umem);
-		mem->umem = NULL;
-	}
+err_free_mem:
+	erdma_mem_free(dev, mem);
 
 	return ret;
 }
 
-static void put_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem)
+static void erdma_mem_uninit(struct erdma_dev *dev, struct erdma_mem *mem)
 {
 	if (mem->mtt) {
 		erdma_destroy_mtt(dev, mem->mtt);
 		mem->mtt = NULL;
 	}
 
-	if (mem->type == ERDMA_KMEM) {
-		erdma_free_kmem(dev, mem);
-	} else if (mem->umem) {
-		ib_umem_release(mem->umem);
-		mem->umem = NULL;
-	}
+	erdma_mem_free(dev, mem);
 }
 
 static int erdma_map_user_dbrecords(struct erdma_ucontext *ctx,
@@ -969,6 +1016,9 @@ erdma_unmap_user_dbrecords(struct erdma_ucontext *ctx,
 static int init_user_qp(struct erdma_qp *qp, struct erdma_ucontext *uctx,
 			u64 va, u32 len, u64 dbrec_va)
 {
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_UMEM,
+	};
 	dma_addr_t dbrec_dma;
 	u32 rq_offset;
 	int ret;
@@ -977,45 +1027,48 @@ static int init_user_qp(struct erdma_qp *qp, struct erdma_ucontext *uctx,
 		   qp->attrs.rq_size * RQE_SIZE))
 		return -EINVAL;
 
-	ret = get_mtt_entries(qp->dev, &qp->user_qp.sq_mem, va,
-			      qp->attrs.sq_size << SQEBB_SHIFT, 0, va,
-			      (SZ_1M - SZ_4K), true);
+	attr.start = va;
+	attr.virt = va;
+	attr.len = qp->attrs.sq_size << SQEBB_SHIFT;
+	attr.req_page_size = SZ_1M - SZ_4K;
+	ret = erdma_mem_init(qp->dev, &qp->user_qp.sq_mem, &attr);
 	if (ret)
 		return ret;
 
 	rq_offset = ALIGN(qp->attrs.sq_size << SQEBB_SHIFT, ERDMA_HW_PAGE_SIZE);
 	qp->user_qp.rq_offset = rq_offset;
 
-	ret = get_mtt_entries(qp->dev, &qp->user_qp.rq_mem, va + rq_offset,
-			      qp->attrs.rq_size << RQE_SHIFT, 0, va + rq_offset,
-			      (SZ_1M - SZ_4K), true);
+	attr.start = va + rq_offset;
+	attr.virt = va + rq_offset;
+	attr.len = qp->attrs.rq_size << RQE_SHIFT;
+	ret = erdma_mem_init(qp->dev, &qp->user_qp.rq_mem, &attr);
 	if (ret)
-		goto put_sq_mtt;
+		goto uninit_sq_mem;
 
 	ret = erdma_map_user_dbrecords(uctx, dbrec_va,
 				       &qp->user_qp.user_dbr_page,
 				       &dbrec_dma);
 	if (ret)
-		goto put_rq_mtt;
+		goto uninit_rq_mem;
 
 	qp->user_qp.sq_dbrec_dma = dbrec_dma;
 	qp->user_qp.rq_dbrec_dma = dbrec_dma + ERDMA_DB_SIZE;
 
 	return 0;
 
-put_rq_mtt:
-	put_mtt_entries(qp->dev, &qp->user_qp.rq_mem);
+uninit_rq_mem:
+	erdma_mem_uninit(qp->dev, &qp->user_qp.rq_mem);
 
-put_sq_mtt:
-	put_mtt_entries(qp->dev, &qp->user_qp.sq_mem);
+uninit_sq_mem:
+	erdma_mem_uninit(qp->dev, &qp->user_qp.sq_mem);
 
 	return ret;
 }
 
 static void free_user_qp(struct erdma_qp *qp, struct erdma_ucontext *uctx)
 {
-	put_mtt_entries(qp->dev, &qp->user_qp.sq_mem);
-	put_mtt_entries(qp->dev, &qp->user_qp.rq_mem);
+	erdma_mem_uninit(qp->dev, &qp->user_qp.sq_mem);
+	erdma_mem_uninit(qp->dev, &qp->user_qp.rq_mem);
 	erdma_unmap_user_dbrecords(uctx, &qp->user_qp.user_dbr_page);
 }
 
@@ -1262,8 +1315,11 @@ struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 				u64 virt, int access, struct ib_dmah *dmah,
 				struct ib_udata *udata)
 {
-	struct erdma_mr *mr = NULL;
 	struct erdma_dev *dev = to_edev(ibpd->device);
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_UMEM,
+	};
+	struct erdma_mr *mr = NULL;
 	u32 stag;
 	int ret;
 
@@ -1277,8 +1333,13 @@ struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 	if (!mr)
 		return ERR_PTR(-ENOMEM);
 
-	ret = get_mtt_entries(dev, &mr->mem, start, len, access, virt,
-			      SZ_2G - SZ_4K, false);
+	attr.start = start;
+	attr.virt = virt;
+	attr.len = len;
+	attr.req_page_size = SZ_2G - SZ_4K;
+	attr.access = access;
+	attr.flags = ERDMA_MEM_FLAG_MR_BUF;
+	ret = erdma_mem_init(dev, &mr->mem, &attr);
 	if (ret)
 		goto err_out_free;
 
@@ -1288,8 +1349,6 @@ struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 
 	mr->ibmr.lkey = mr->ibmr.rkey = stag;
 	mr->ibmr.pd = ibpd;
-	mr->mem.va = virt;
-	mr->mem.len = len;
 	mr->access = ERDMA_MR_ACC_LR | to_erdma_access_flags(access);
 	mr->valid = 1;
 	mr->type = ERDMA_MR_TYPE_NORMAL;
@@ -1305,7 +1364,7 @@ struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 		       mr->ibmr.lkey >> 8);
 
 err_out_put_mtt:
-	put_mtt_entries(dev, &mr->mem);
+	erdma_mem_uninit(dev, &mr->mem);
 
 err_out_free:
 	kfree(mr);
@@ -1342,7 +1401,7 @@ int erdma_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 
 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX], ibmr->lkey >> 8);
 
-	put_mtt_entries(dev, &mr->mem);
+	erdma_mem_uninit(dev, &mr->mem);
 
 	kfree(mr);
 	return 0;
@@ -1377,12 +1436,12 @@ int erdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 	wait_for_completion(&cq->free);
 
 	if (rdma_is_kernel_res(&cq->ibcq.res)) {
-		put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
+		erdma_mem_uninit(dev, &cq->kern_cq.qbuf_mem);
 		dma_pool_free(dev->db_pool, cq->kern_cq.dbrec,
 			      cq->kern_cq.dbrec_dma);
 	} else {
 		erdma_unmap_user_dbrecords(ctx, &cq->user_cq.user_dbr_page);
-		put_mtt_entries(dev, &cq->user_cq.qbuf_mem);
+		erdma_mem_uninit(dev, &cq->user_cq.qbuf_mem);
 	}
 
 	return 0;
@@ -1434,8 +1493,8 @@ int erdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 	if (rdma_is_kernel_res(&qp->ibqp.res)) {
 		free_kernel_qp(qp);
 	} else {
-		put_mtt_entries(dev, &qp->user_qp.sq_mem);
-		put_mtt_entries(dev, &qp->user_qp.rq_mem);
+		erdma_mem_uninit(dev, &qp->user_qp.sq_mem);
+		erdma_mem_uninit(dev, &qp->user_qp.rq_mem);
 		erdma_unmap_user_dbrecords(ctx, &qp->user_qp.user_dbr_page);
 	}
 
@@ -1953,12 +2012,17 @@ int erdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 static int erdma_init_user_cq(struct erdma_ucontext *ctx, struct erdma_cq *cq,
 			      struct erdma_ureq_create_cq *ureq)
 {
-	int ret;
 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_UMEM,
+	};
+	int ret;
 
-	ret = get_mtt_entries(dev, &cq->user_cq.qbuf_mem, ureq->qbuf_va,
-			      ureq->qbuf_len, 0, ureq->qbuf_va, SZ_64M - SZ_4K,
-			      true);
+	attr.start = ureq->qbuf_va;
+	attr.virt = ureq->qbuf_va;
+	attr.len = ureq->qbuf_len;
+	attr.req_page_size = SZ_64M - SZ_4K;
+	ret = erdma_mem_init(dev, &cq->user_cq.qbuf_mem, &attr);
 	if (ret)
 		return ret;
 
@@ -1966,7 +2030,7 @@ static int erdma_init_user_cq(struct erdma_ucontext *ctx, struct erdma_cq *cq,
 				       &cq->user_cq.user_dbr_page,
 				       &cq->user_cq.dbrec_dma);
 	if (ret)
-		put_mtt_entries(dev, &cq->user_cq.qbuf_mem);
+		erdma_mem_uninit(dev, &cq->user_cq.qbuf_mem);
 
 	return ret;
 }
@@ -1974,10 +2038,15 @@ static int erdma_init_user_cq(struct erdma_ucontext *ctx, struct erdma_cq *cq,
 static int erdma_init_kernel_cq(struct erdma_cq *cq)
 {
 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_KMEM,
+		.len = cq->depth << CQE_SHIFT,
+	};
+	int ret;
 
-	if (erdma_alloc_kmem(dev, &cq->kern_cq.qbuf_mem,
-			     cq->depth << CQE_SHIFT))
-		return -ENOMEM;
+	ret = erdma_mem_init(dev, &cq->kern_cq.qbuf_mem, &attr);
+	if (ret)
+		return ret;
 
 	cq->kern_cq.dbrec = dma_pool_zalloc(dev->db_pool, GFP_KERNEL,
 					    &cq->kern_cq.dbrec_dma);
@@ -1991,7 +2060,7 @@ static int erdma_init_kernel_cq(struct erdma_cq *cq)
 	return 0;
 
 err_out:
-	put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
+	erdma_mem_uninit(dev, &cq->kern_cq.qbuf_mem);
 
 	return -ENOMEM;
 }
@@ -2056,9 +2125,9 @@ int erdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 err_free_res:
 	if (!rdma_is_kernel_res(&ibcq->res)) {
 		erdma_unmap_user_dbrecords(ctx, &cq->user_cq.user_dbr_page);
-		put_mtt_entries(dev, &cq->user_cq.qbuf_mem);
+		erdma_mem_uninit(dev, &cq->user_cq.qbuf_mem);
 	} else {
-		put_mtt_entries(dev, &cq->kern_cq.qbuf_mem);
+		erdma_mem_uninit(dev, &cq->kern_cq.qbuf_mem);
 		dma_pool_free(dev->db_pool, cq->kern_cq.dbrec,
 			      cq->kern_cq.dbrec_dma);
 	}
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.h b/drivers/infiniband/hw/erdma/erdma_verbs.h
index 5056d60357db..4e72b3c22aad 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.h
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.h
@@ -125,19 +125,34 @@ struct erdma_kmem {
 	struct erdma_buf_list *buf_list;
 };
 
+enum erdma_mem_flags {
+	ERDMA_MEM_FLAG_MR_BUF = (1 << 0),
+};
+
+struct erdma_mem_init_attr {
+	enum erdma_mem_type type;
+	u64 start;
+	u64 virt;
+	u64 len;
+	unsigned long req_page_size;
+	int access;
+	u32 flags;
+};
+
 struct erdma_mem {
 	enum erdma_mem_type type;
 	union {
 		struct ib_umem *umem;
 		struct erdma_kmem *kmem;
 	};
-	struct erdma_mtt *mtt;
 
 	u32 page_size;
 	u32 page_offset;
 	u32 page_cnt;
 	u32 mtt_nents;
 
+	struct erdma_mtt *mtt;
+
 	u64 va;
 	u64 len;
 };
-- 
2.31.1

[PATCH for-next v2 4/4] RDMA/erdma: Move kernel QP helpers after memory helpers

[Cheng Xu]

Move the kernel QP helpers after the shared memory helpers to remove
forward declarations.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma_verbs.c | 166 +++++++++++-----------
 1 file changed, 81 insertions(+), 85 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index aaef2b3621b0..482a3d603fe0 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -19,10 +19,6 @@
 #include "erdma_cm.h"
 #include "erdma_verbs.h"
 
-static int erdma_mem_init(struct erdma_dev *dev, struct erdma_mem *mem,
-			  struct erdma_mem_init_attr *attr);
-static void erdma_mem_uninit(struct erdma_dev *dev, struct erdma_mem *mem);
-
 static void assemble_qbuf_mtt_for_cmd(struct erdma_mem *mem, u32 *cfg,
 				      u64 *addr0, u64 *addr1)
 {
@@ -473,87 +469,6 @@ static int erdma_qp_validate_attr(struct erdma_dev *dev,
 	return 0;
 }
 
-static void free_kernel_qp(struct erdma_qp *qp)
-{
-	struct erdma_dev *dev = qp->dev;
-
-	vfree(qp->kern_qp.swr_tbl);
-	vfree(qp->kern_qp.rwr_tbl);
-
-	erdma_mem_uninit(dev, &qp->kern_qp.sq_mem);
-
-	if (qp->kern_qp.sq_dbrec)
-		dma_pool_free(dev->db_pool, qp->kern_qp.sq_dbrec,
-			      qp->kern_qp.sq_dbrec_dma);
-
-	erdma_mem_uninit(dev, &qp->kern_qp.rq_mem);
-
-	if (qp->kern_qp.rq_dbrec)
-		dma_pool_free(dev->db_pool, qp->kern_qp.rq_dbrec,
-			      qp->kern_qp.rq_dbrec_dma);
-}
-
-static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
-			  struct ib_qp_init_attr *attrs)
-{
-	struct erdma_mem_init_attr attr = {
-		.type = ERDMA_KMEM,
-	};
-	struct erdma_kqp *kqp = &qp->kern_qp;
-	int size;
-	int ret;
-
-	if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
-		kqp->sig_all = 1;
-
-	kqp->sq_pi = 0;
-	kqp->sq_ci = 0;
-	kqp->rq_pi = 0;
-	kqp->rq_ci = 0;
-	kqp->hw_sq_db =
-		dev->func_bar + (ERDMA_SDB_SHARED_PAGE_INDEX << PAGE_SHIFT);
-	kqp->hw_rq_db = dev->func_bar + ERDMA_BAR_RQDB_SPACE_OFFSET;
-
-	kqp->swr_tbl = vmalloc_array(qp->attrs.sq_size, sizeof(u64));
-	kqp->rwr_tbl = vmalloc_array(qp->attrs.rq_size, sizeof(u64));
-	if (!kqp->swr_tbl || !kqp->rwr_tbl) {
-		ret = -ENOMEM;
-		goto err_out;
-	}
-
-	size = qp->attrs.sq_size << SQEBB_SHIFT;
-	attr.len = size;
-	ret = erdma_mem_init(dev, &kqp->sq_mem, &attr);
-	if (ret)
-		goto err_out;
-
-	kqp->sq_dbrec =
-		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->sq_dbrec_dma);
-	if (!kqp->sq_dbrec) {
-		ret = -ENOMEM;
-		goto err_out;
-	}
-
-	size = qp->attrs.rq_size << RQE_SHIFT;
-	attr.len = size;
-	ret = erdma_mem_init(dev, &kqp->rq_mem, &attr);
-	if (ret)
-		goto err_out;
-
-	kqp->rq_dbrec =
-		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->rq_dbrec_dma);
-	if (!kqp->rq_dbrec) {
-		ret = -ENOMEM;
-		goto err_out;
-	}
-
-	return 0;
-
-err_out:
-	free_kernel_qp(qp);
-	return ret;
-}
-
 static void erdma_init_mtt_leaf(struct erdma_mem *mem, struct erdma_mtt *mtt)
 {
 	struct ib_block_iter biter;
@@ -951,6 +866,87 @@ static void erdma_mem_uninit(struct erdma_dev *dev, struct erdma_mem *mem)
 	erdma_mem_free(dev, mem);
 }
 
+static void free_kernel_qp(struct erdma_qp *qp)
+{
+	struct erdma_dev *dev = qp->dev;
+
+	vfree(qp->kern_qp.swr_tbl);
+	vfree(qp->kern_qp.rwr_tbl);
+
+	erdma_mem_uninit(dev, &qp->kern_qp.sq_mem);
+
+	if (qp->kern_qp.sq_dbrec)
+		dma_pool_free(dev->db_pool, qp->kern_qp.sq_dbrec,
+			      qp->kern_qp.sq_dbrec_dma);
+
+	erdma_mem_uninit(dev, &qp->kern_qp.rq_mem);
+
+	if (qp->kern_qp.rq_dbrec)
+		dma_pool_free(dev->db_pool, qp->kern_qp.rq_dbrec,
+			      qp->kern_qp.rq_dbrec_dma);
+}
+
+static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
+			  struct ib_qp_init_attr *attrs)
+{
+	struct erdma_mem_init_attr attr = {
+		.type = ERDMA_KMEM,
+	};
+	struct erdma_kqp *kqp = &qp->kern_qp;
+	int size;
+	int ret;
+
+	if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
+		kqp->sig_all = 1;
+
+	kqp->sq_pi = 0;
+	kqp->sq_ci = 0;
+	kqp->rq_pi = 0;
+	kqp->rq_ci = 0;
+	kqp->hw_sq_db =
+		dev->func_bar + (ERDMA_SDB_SHARED_PAGE_INDEX << PAGE_SHIFT);
+	kqp->hw_rq_db = dev->func_bar + ERDMA_BAR_RQDB_SPACE_OFFSET;
+
+	kqp->swr_tbl = vmalloc_array(qp->attrs.sq_size, sizeof(u64));
+	kqp->rwr_tbl = vmalloc_array(qp->attrs.rq_size, sizeof(u64));
+	if (!kqp->swr_tbl || !kqp->rwr_tbl) {
+		ret = -ENOMEM;
+		goto err_out;
+	}
+
+	size = qp->attrs.sq_size << SQEBB_SHIFT;
+	attr.len = size;
+	ret = erdma_mem_init(dev, &kqp->sq_mem, &attr);
+	if (ret)
+		goto err_out;
+
+	kqp->sq_dbrec =
+		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->sq_dbrec_dma);
+	if (!kqp->sq_dbrec) {
+		ret = -ENOMEM;
+		goto err_out;
+	}
+
+	size = qp->attrs.rq_size << RQE_SHIFT;
+	attr.len = size;
+	ret = erdma_mem_init(dev, &kqp->rq_mem, &attr);
+	if (ret)
+		goto err_out;
+
+	kqp->rq_dbrec =
+		dma_pool_zalloc(dev->db_pool, GFP_KERNEL, &kqp->rq_dbrec_dma);
+	if (!kqp->rq_dbrec) {
+		ret = -ENOMEM;
+		goto err_out;
+	}
+
+	return 0;
+
+err_out:
+	free_kernel_qp(qp);
+	return ret;
+}
+
 static int erdma_map_user_dbrecords(struct erdma_ucontext *ctx,
 				    u64 dbrecords_va,
 				    struct erdma_user_dbrecords_page **dbr_page,
-- 
2.31.1

Re: [PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Leon Romanovsky]

On Thu, Aug 27, 2026 at 04:25:20PM +0800, Cheng Xu wrote:
> A single coherent allocation for kernel QP queues can fail for large
> queues when memory is fragmented.
> 
> Allocate page-sized coherent buffers and describe them with the existing
> MTT. Keep the userspace QP path unchanged.
> 
> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
> ---
>  drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
>  drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
>  drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
>  drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
>  4 files changed, 179 insertions(+), 102 deletions(-)

<...>

> +static void erdma_free_kmem(struct erdma_dev *dev, struct erdma_mem *mem)
> +{
> +	struct erdma_buf_list *buf_list;
> +	u32 i;
> +
> +	if (!mem->kmem)
> +		return;

Can it be null?

> +
> +	buf_list = mem->kmem->buf_list;
> +	if (buf_list) {

Write the unwinding section of erdma_alloc_kmem() to don't call to
erdma_free_kmem().

> +		for (i = 0; i < mem->page_cnt; i++)
> +			if (buf_list[i].buf)
> +				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
> +						  buf_list[i].buf,
> +						  buf_list[i].dma_addr);
> +		kfree(buf_list);
> +	}
> +
> +	kfree(mem->kmem);
> +	mem->kmem = NULL;
> +}
> +
> +static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
> +			    size_t size)
> +{
> +	struct erdma_buf_list *buf_list;
> +	u32 i;
> +
> +	mem->type = ERDMA_KMEM;
> +	mem->page_size = PAGE_SIZE;
> +	mem->page_offset = 0;
> +	mem->page_cnt = DIV_ROUND_UP(size, PAGE_SIZE);
> +	mem->mtt_nents = mem->page_cnt;
> +	mem->len = size;
> +
> +	mem->kmem = kzalloc_obj(*mem->kmem);
> +	if (!mem->kmem)
> +		goto err_free_mem;

There is no point to call to erdma_free_kmem() here and you can return
immediately.

> +
> +	buf_list = kcalloc(mem->page_cnt, sizeof(*buf_list), GFP_KERNEL);
> +	if (!buf_list)
> +		goto err_free_mem;

Please write error unwinding explicitly.

> +	mem->kmem->buf_list = buf_list;
> +
> +	for (i = 0; i < mem->page_cnt; i++) {
> +		buf_list[i].buf = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
> +						     &buf_list[i].dma_addr,
> +						     GFP_KERNEL);
> +		if (!buf_list[i].buf)
> +			goto err_free_mem;
> +	}
> +
> +	mem->mtt = erdma_create_mtt(dev, MTT_SIZE(mem->page_cnt), true);
> +	if (IS_ERR(mem->mtt)) {
> +		mem->mtt = NULL;

Why? You will release mem anyway.

> +		goto err_free_mem;
> +	}
> +	for (i = 0; i < mem->page_cnt; i++)
> +		mem->mtt->buf[i] = buf_list[i].dma_addr;
> +
> +	return 0;
> +
> +err_free_mem:
> +	erdma_free_kmem(dev, mem);
> +	return -ENOMEM;
> +}
> +

<...>

> +struct erdma_buf_list {
> +	void *buf;
> +	dma_addr_t dma_addr;
> +};

This struct is very similar to scatter-gather list, why don't you use it
directly?

Thanks

Re: [PATCH for-next v2 3/4] RDMA/erdma: Unify userspace and kernel queue buffer management

[Leon Romanovsky]

On Thu, Aug 27, 2026 at 04:25:22PM +0800, Cheng Xu wrote:
> Userspace and kernel space queue buffers use separate helpers despite
> sharing MTT metadata and lifetime rules. Manage both through
> erdma_mem_init() and erdma_mem_uninit(), while keeping backing allocation
> and release type-specific.
> 
> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
> ---
>  drivers/infiniband/hw/erdma/erdma_verbs.c | 281 ++++++++++++++--------
>  drivers/infiniband/hw/erdma/erdma_verbs.h |  17 +-
>  2 files changed, 191 insertions(+), 107 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
> index b0a47c82cbdf..aaef2b3621b0 100644
> --- a/drivers/infiniband/hw/erdma/erdma_verbs.c
> +++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
> @@ -19,9 +19,9 @@
>  #include "erdma_cm.h"
>  #include "erdma_verbs.h"

<...>

> -static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
> -			   u64 start, u64 len, int access, u64 virt,
> -			   unsigned long req_page_size, bool force_continuous)
> +static void erdma_mem_free(struct erdma_dev *dev, struct erdma_mem *mem)
>  {
> -	int ret = 0;
> +	switch (mem->type) {
> +	case ERDMA_UMEM:
> +		if (mem->umem) {

I expect that if you call to this function, mem->umem is already valid.

> +			ib_umem_release(mem->umem);
> +			mem->umem = NULL;
> +		}
> +		break;
> +	case ERDMA_KMEM:
> +		erdma_kmem_free(dev, mem);
> +		break;
> +	default:
> +		break;
> +	}
> +}
>  

Thanks

Re: [PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Cheng Xu]

On 9/3/26 5:10 PM, Leon Romanovsky wrote:
> On Thu, Aug 27, 2026 at 04:25:20PM +0800, Cheng Xu wrote:
>> A single coherent allocation for kernel QP queues can fail for large
>> queues when memory is fragmented.
>>
>> Allocate page-sized coherent buffers and describe them with the existing
>> MTT. Keep the userspace QP path unchanged.
>>
>> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
>> ---
>>  drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
>>  drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
>>  drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
>>  drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
>>  4 files changed, 179 insertions(+), 102 deletions(-)
> 
> <...>
> 
>> +static void erdma_free_kmem(struct erdma_dev *dev, struct erdma_mem *mem)
>> +{
>> +	struct erdma_buf_list *buf_list;
>> +	u32 i;
>> +
>> +	if (!mem->kmem)
>> +		return;
> 
> Can it be null?

Yes. With the current init_kernel_qp() implementation, failures go
through the common free_kernel_qp() path instead of unwinding each
completed allocation in reverse order. As a result, erdma_free_kmem()
may be called more than once for partially initialized memory, and
mem->kmem can be NULL.

This is not ideal. I will add a separate preparatory patch to implement
explicit error unwinding in init_kernel_qp().

> 
>> +
>> +	buf_list = mem->kmem->buf_list;
>> +	if (buf_list) {
> 
> Write the unwinding section of erdma_alloc_kmem() to don't call to
> erdma_free_kmem().

Agreed. Patch #3 already refactors erdma_alloc_kmem() to unwind its
allocations explicitly instead of calling erdma_free_kmem(). However,
each intermediate patch should have correct error handling, so I will
move the change into this patch.

> 
>> +		for (i = 0; i < mem->page_cnt; i++)
>> +			if (buf_list[i].buf)
>> +				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
>> +						  buf_list[i].buf,
>> +						  buf_list[i].dma_addr);
>> +		kfree(buf_list);
>> +	}
>> +
>> +	kfree(mem->kmem);
>> +	mem->kmem = NULL;
>> +}
>> +
>> +static int erdma_alloc_kmem(struct erdma_dev *dev, struct erdma_mem *mem,
>> +			    size_t size)
>> +{
>> +	struct erdma_buf_list *buf_list;
>> +	u32 i;
>> +
>> +	mem->type = ERDMA_KMEM;
>> +	mem->page_size = PAGE_SIZE;
>> +	mem->page_offset = 0;
>> +	mem->page_cnt = DIV_ROUND_UP(size, PAGE_SIZE);
>> +	mem->mtt_nents = mem->page_cnt;
>> +	mem->len = size;
>> +
>> +	mem->kmem = kzalloc_obj(*mem->kmem);
>> +	if (!mem->kmem)
>> +		goto err_free_mem;
> 
> There is no point to call to erdma_free_kmem() here and you can return
> immediately.
> 

Right. No kmem resource has been allocated at this point, so it should
return -ENOMEM directly. This is already handled in patch #3, and I
will move the change into this patch.

>> +
>> +	buf_list = kcalloc(mem->page_cnt, sizeof(*buf_list), GFP_KERNEL);
>> +	if (!buf_list)
>> +		goto err_free_mem;
> 
> Please write error unwinding explicitly.

Likewise. The explicit error-unwinding paths are already introduced in
patch #3. I will move them into this patch.


> 
>> +	mem->kmem->buf_list = buf_list;
>> +
>> +	for (i = 0; i < mem->page_cnt; i++) {
>> +		buf_list[i].buf = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
>> +						     &buf_list[i].dma_addr,
>> +						     GFP_KERNEL);
>> +		if (!buf_list[i].buf)
>> +			goto err_free_mem;
>> +	}
>> +
>> +	mem->mtt = erdma_create_mtt(dev, MTT_SIZE(mem->page_cnt), true);
>> +	if (IS_ERR(mem->mtt)) {
>> +		mem->mtt = NULL;
> 
> Why? You will release mem anyway.
> 

This was needed because the current code lacks explicit error
unwinding. The common cleanup path only checks mem->mtt for NULL and
would otherwise pass the ERR_PTR to erdma_destroy_mtt().

This has the same root cause as the first issue, and I will fix it in
this patch as part of the same error-unwinding change.


>> +		goto err_free_mem;
>> +	}
>> +	for (i = 0; i < mem->page_cnt; i++)
>> +		mem->mtt->buf[i] = buf_list[i].dma_addr;
>> +
>> +	return 0;
>> +
>> +err_free_mem:
>> +	erdma_free_kmem(dev, mem);
>> +	return -ENOMEM;
>> +}
>> +
> 
> <...>
> 
>> +struct erdma_buf_list {
>> +	void *buf;
>> +	dma_addr_t dma_addr;
>> +};
> 
> This struct is very similar to scatter-gather list, why don't you use it
> directly?

Good idea. I will use struct scatterlist in the next revision.

Thanks,
Cheng Xu

> 
> Thanks

Re: [PATCH for-next v2 3/4] RDMA/erdma: Unify userspace and kernel queue buffer management

[Cheng Xu]

On 9/3/26 5:12 PM, Leon Romanovsky wrote:
> On Thu, Aug 27, 2026 at 04:25:22PM +0800, Cheng Xu wrote:
>> Userspace and kernel space queue buffers use separate helpers despite
>> sharing MTT metadata and lifetime rules. Manage both through
>> erdma_mem_init() and erdma_mem_uninit(), while keeping backing allocation
>> and release type-specific.
>>
>> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
>> ---
>>  drivers/infiniband/hw/erdma/erdma_verbs.c | 281 ++++++++++++++--------
>>  drivers/infiniband/hw/erdma/erdma_verbs.h |  17 +-
>>  2 files changed, 191 insertions(+), 107 deletions(-)
>>
>> diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
>> index b0a47c82cbdf..aaef2b3621b0 100644
>> --- a/drivers/infiniband/hw/erdma/erdma_verbs.c
>> +++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
>> @@ -19,9 +19,9 @@
>>  #include "erdma_cm.h"
>>  #include "erdma_verbs.h"
> 
> <...>
> 
>> -static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
>> -			   u64 start, u64 len, int access, u64 virt,
>> -			   unsigned long req_page_size, bool force_continuous)
>> +static void erdma_mem_free(struct erdma_dev *dev, struct erdma_mem *mem)
>>  {
>> -	int ret = 0;
>> +	switch (mem->type) {
>> +	case ERDMA_UMEM:
>> +		if (mem->umem) {
> 
> I expect that if you call to this function, mem->umem is already valid.
> 

Sure, I will fix it.

Thanks,
Cheng Xu

>> +			ib_umem_release(mem->umem);
>> +			mem->umem = NULL;
>> +		}
>> +		break;
>> +	case ERDMA_KMEM:
>> +		erdma_kmem_free(dev, mem);
>> +		break;
>> +	default:
>> +		break;
>> +	}
>> +}
>>  
> 
> Thanks

Re: [PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Cheng Xu]

On 9/3/26 8:38 PM, Cheng Xu wrote:
> 
> 
> On 9/3/26 5:10 PM, Leon Romanovsky wrote:
>> On Thu, Aug 27, 2026 at 04:25:20PM +0800, Cheng Xu wrote:
>>> A single coherent allocation for kernel QP queues can fail for large
>>> queues when memory is fragmented.
>>>
>>> Allocate page-sized coherent buffers and describe them with the existing
>>> MTT. Keep the userspace QP path unchanged.
>>>
>>> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
>>> ---
>>>  drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
>>>  drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
>>>  drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
>>>  drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
>>>  4 files changed, 179 insertions(+), 102 deletions(-)
>>
>> <...>
>>

<...>

>>
>>> +struct erdma_buf_list {
>>> +	void *buf;
>>> +	dma_addr_t dma_addr;
>>> +};
>>
>> This struct is very similar to scatter-gather list, why don't you use it
>> directly?
> 
> Good idea. I will use struct scatterlist in the next revision.

Hi Leon,

I switched to scatterlist in v3, but Sashiko pointed out an issue with
using sg_set_buf() and sg_virt() on dma_alloc_coherent() memory [1].

To handle this correctly, the driver would still need to retain the
original CPU addresses returned by dma_alloc_coherent(). Using scatterlist
does not simplify this implementation: we still need separate storage for
the CPU addresses, and the only scatterlist field we actually need is the
DMA address. A small structure holding both addresses would therefore be
simpler.

I would like to switch back to a small structure holding the CPU and DMA
addresses in v4. This would keep the implementation simpler while
preserving the original addresses returned by dma_alloc_coherent().

[1] https://sashiko.dev/#/patchset/20260908033709.89898-1-chengyou%40linux.alibaba.com

Thanks,
Cheng Xu


> Thanks,
> Cheng Xu
> 
>>
>> Thanks

Re: [PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Leon Romanovsky]

On Wed, Sep 09, 2026 at 02:21:37PM +0800, Cheng Xu wrote:
> 
> 
> On 9/3/26 8:38 PM, Cheng Xu wrote:
> > 
> > 
> > On 9/3/26 5:10 PM, Leon Romanovsky wrote:
> >> On Thu, Aug 27, 2026 at 04:25:20PM +0800, Cheng Xu wrote:
> >>> A single coherent allocation for kernel QP queues can fail for large
> >>> queues when memory is fragmented.
> >>>
> >>> Allocate page-sized coherent buffers and describe them with the existing
> >>> MTT. Keep the userspace QP path unchanged.
> >>>
> >>> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
> >>> ---
> >>>  drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
> >>>  drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
> >>>  drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
> >>>  drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
> >>>  4 files changed, 179 insertions(+), 102 deletions(-)
> >>
> >> <...>
> >>
> 
> <...>
> 
> >>
> >>> +struct erdma_buf_list {
> >>> +	void *buf;
> >>> +	dma_addr_t dma_addr;
> >>> +};
> >>
> >> This struct is very similar to scatter-gather list, why don't you use it
> >> directly?
> > 
> > Good idea. I will use struct scatterlist in the next revision.
> 
> Hi Leon,
> 
> I switched to scatterlist in v3, but Sashiko pointed out an issue with
> using sg_set_buf() and sg_virt() on dma_alloc_coherent() memory [1].
> 
> To handle this correctly, the driver would still need to retain the
> original CPU addresses returned by dma_alloc_coherent(). Using scatterlist
> does not simplify this implementation: we still need separate storage for
> the CPU addresses, and the only scatterlist field we actually need is the
> DMA address. A small structure holding both addresses would therefore be
> simpler.

Sashiko thinks that you are creating SG list to feed it to dma_map_sg()
later which is not. You are using SG as simple database and you will get
iterators for free.

cpu_address = sg_page()
dma_address = sg_dma_address()

Thanks

> 
> I would like to switch back to a small structure holding the CPU and DMA
> addresses in v4. This would keep the implementation simpler while
> preserving the original addresses returned by dma_alloc_coherent().
> 
> [1] https://sashiko.dev/#/patchset/20260908033709.89898-1-chengyou%40linux.alibaba.com
> 
> Thanks,
> Cheng Xu
> 
> 
> > Thanks,
> > Cheng Xu
> > 
> >>
> >> Thanks

Re: [PATCH for-next v2 1/4] RDMA/erdma: Support non-contiguous kernel QP buffers

[Cheng Xu]

On 9/10/26 11:33 PM, Leon Romanovsky wrote:
> On Wed, Sep 09, 2026 at 02:21:37PM +0800, Cheng Xu wrote:
>>
>>
>> On 9/3/26 8:38 PM, Cheng Xu wrote:
>>>
>>>
>>> On 9/3/26 5:10 PM, Leon Romanovsky wrote:
>>>> On Thu, Aug 27, 2026 at 04:25:20PM +0800, Cheng Xu wrote:
>>>>> A single coherent allocation for kernel QP queues can fail for large
>>>>> queues when memory is fragmented.
>>>>>
>>>>> Allocate page-sized coherent buffers and describe them with the existing
>>>>> MTT. Keep the userspace QP path unchanged.
>>>>>
>>>>> Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
>>>>> ---
>>>>>  drivers/infiniband/hw/erdma/erdma_cq.c    |   4 +-
>>>>>  drivers/infiniband/hw/erdma/erdma_qp.c    |  38 +++--
>>>>>  drivers/infiniband/hw/erdma/erdma_verbs.c | 195 +++++++++++++---------
>>>>>  drivers/infiniband/hw/erdma/erdma_verbs.h |  44 ++++-
>>>>>  4 files changed, 179 insertions(+), 102 deletions(-)
>>>>
>>>> <...>
>>>>
>>
>> <...>
>>
>>>>
>>>>> +struct erdma_buf_list {
>>>>> +	void *buf;
>>>>> +	dma_addr_t dma_addr;
>>>>> +};
>>>>
>>>> This struct is very similar to scatter-gather list, why don't you use it
>>>> directly?
>>>
>>> Good idea. I will use struct scatterlist in the next revision.
>>
>> Hi Leon,
>>
>> I switched to scatterlist in v3, but Sashiko pointed out an issue with
>> using sg_set_buf() and sg_virt() on dma_alloc_coherent() memory [1].
>>
>> To handle this correctly, the driver would still need to retain the
>> original CPU addresses returned by dma_alloc_coherent(). Using scatterlist
>> does not simplify this implementation: we still need separate storage for
>> the CPU addresses, and the only scatterlist field we actually need is the
>> DMA address. A small structure holding both addresses would therefore be
>> simpler.
> 
> Sashiko thinks that you are creating SG list to feed it to dma_map_sg()
> later which is not. You are using SG as simple database and you will get
> iterators for free.
> 
> cpu_address = sg_page()
> dma_address = sg_dma_address()
> 

Hi Leon,

Maybe I misunderstood something. My understanding of Sashiko's concern is
that the CPU address returned by dma_alloc_coherent() is not guaranteed to
be valid for virt_to_page(), as used by sg_set_buf(). Consequently, the
struct page later returned by sg_page() may be invalid.

Although this works on the x86 platforms we tested, it may not be portable
to architectures where coherent memory is remapped outside the linear
mapping. From this perspective, would a small structure that explicitly
stores the CPU and DMA addresses, similar to those used by HNS and mlx5, be
more appropriate than scatterlist here?

I collected some information related to this below [1][2][3].

[1] include/linux/scatterlist.h:

#ifdef CONFIG_DEBUG_SG
	BUG_ON(!virt_addr_valid(buf));
#endif
sg_set_page(sg, virt_to_page(buf), buflen, offset_in_page(buf));

[2] kernel/dma/mapping.c:

/*
 * The whole dma_get_sgtable() idea is fundamentally unsafe ...
 * 1. Not all memory allocated via the coherent DMA APIs is backed by
 *    a struct page
 */

[3] drivers/infiniband/hw/mthca/mthca_memfree.c:

/* We use sg_set_buf for coherent allocs, which assumes low memory */

Thanks,
Cheng Xu


> Thanks
>