Skip to contentSkip to conversation

Implement memory windows support for rxe

PATCH In review Bob Pearson 2 participants 9 messages Jun 8, 2021 – Jun 9, 2021 InfiniBand

lore
Share & tools
lore
#0 lore raw

v2 cover Implement memory windows support for rxe

These patches makes the required changes to the rxe provider and the rxe ABI to support the kernel memory windows patches to the rxe driver.

---

Bob Pearson (2):
Update kernel headers
Providers/rxe: Implement memory windows

 kernel-headers/rdma/rdma_user_rxe.h |  10 ++
 providers/rxe/rxe.c                 | 157 +++++++++++++++++++++++++++-
 2 files changed, 165 insertions(+), 2 deletions(-)
Signature · 1 line
2.30.2
#1 lore raw

PATCH v2 1/2 Update kernel headers

To commit ?? ("RDMA/rxe: Disallow MR dereg and invalidate when bound").

---
 kernel-headers/rdma/rdma_user_rxe.h | 10 ++++++++++
 1 file changed, 10 insertions(+)
kernel-headers/rdma/rdma_user_rxe.h +10 -0
index 068433e2..e283c222 100644
@@ -99,7 +99,16 @@ struct rxe_send_wr {
 			__u32	remote_qkey;
 			__u16	pkey_index;
 		} ud;
+		struct {
+			__aligned_u64	addr;
+			__aligned_u64	length;
+			__u32		mr_lkey;
+			__u32		mw_rkey;
+			__u32		rkey;
+			__u32		access;
+		} mw;
 		/* reg is only used by the kernel and is not part of the uapi */
+#ifdef __KERNEL__
 		struct {
 			union {
 				struct ib_mr *mr;
@@ -108,6 +117,7 @@ struct rxe_send_wr {
 			__u32	     key;
 			__u32	     access;
 		} reg;
+#endif
 	} wr;
 };
 
Signature · 1 line
2.30.2
#2 lore raw

PATCH v2 2/2 Providers/rxe: Implement memory windows

In reply to Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

---
v2:
Added support for extended QP bind MW work requests.

---
 providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -2
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
 	return ret;
 }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+	int ret;
+	struct ibv_mw *ibmw;
+	struct ibv_alloc_mw cmd = {};
+	struct ib_uverbs_alloc_mw_resp resp = {};
+
+	ibmw = calloc(1, sizeof(*ibmw));
+	if (!ibmw)
+		return NULL;
+
+	ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+						&resp, sizeof(resp));
+	if (ret) {
+		free(ibmw);
+		return NULL;
+	}
+
+	return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+	int ret;
+
+	ret = ibv_cmd_dealloc_mw(ibmw);
+	if (ret)
+		return ret;
+
+	free(ibmw);
+	return 0;
+}
+
+static int next_rkey(int rkey)
+{
+	return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+			 struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+			struct ibv_mw_bind *mw_bind)
+{
+	int ret;
+	struct ibv_mw_bind_info	*bind_info = &mw_bind->bind_info;
+	struct ibv_send_wr ibwr;
+	struct ibv_send_wr *bad_wr;
+
+	if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+		ret = EINVAL;
+		goto err;
+	}
+
+	if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+		ret = EINVAL;
+		goto err;
+	}
+
+	if (bind_info->mr) {
+		if (ibmw->pd != bind_info->mr->pd) {
+			ret = EPERM;
+			goto err;
+		}
+	}
+
+	memset(&ibwr, 0, sizeof(ibwr));
+
+	ibwr.opcode		= IBV_WR_BIND_MW;
+	ibwr.next		= NULL;
+	ibwr.wr_id		= mw_bind->wr_id;
+	ibwr.send_flags		= mw_bind->send_flags;
+	ibwr.bind_mw.bind_info	= mw_bind->bind_info;
+	ibwr.bind_mw.mw		= ibmw;
+	ibwr.bind_mw.rkey	= next_rkey(ibmw->rkey);
+
+	ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+	if (ret)
+		goto err;
+
+	/* user has to undo this if he gets an error wc */
+	ibmw->rkey = ibwr.bind_mw.rkey;
+
+	return 0;
+err:
+	errno = ret;
+	return errno;
+}
+
 static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
 				 uint64_t hca_va, int access)
 {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
 	advance_qp_cur_index(qp);
 }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+		       uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+	struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+	struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+	if (check_qp_queue_full(qp))
+		return;
+
+	memset(wqe, 0, sizeof(*wqe));
+
+	wqe->wr.wr_id = ibqp->wr_id;
+	wqe->wr.opcode = IBV_WR_BIND_MW;
+	wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+	wqe->wr.wr.mw.addr = info->addr;
+	wqe->wr.wr.mw.length = info->length;
+	wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+	wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+	wqe->wr.wr.mw.rkey = rkey;
+	wqe->wr.wr.mw.access = info->mw_access_flags;
+	wqe->ssn = qp->ssn++;
+
+	advance_qp_cur_index(qp);
+}
+
 static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
 {
 	struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
 		| IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
 		| IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
 		| IBV_QP_EX_WITH_LOCAL_INV
+		| IBV_QP_EX_WITH_BIND_MW
 		| IBV_QP_EX_WITH_SEND_WITH_INV,
 
 	RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
 		| IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
 		| IBV_QP_EX_WITH_SEND
 		| IBV_QP_EX_WITH_SEND_WITH_IMM
+		| IBV_QP_EX_WITH_BIND_MW
 		| IBV_QP_EX_WITH_SEND_WITH_INV,
 
 	RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
 	if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
 		qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+	if (flags & IBV_QP_EX_WITH_BIND_MW)
+		qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
 	if (flags & IBV_QP_EX_WITH_LOCAL_INV)
 		qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
 }
 
 /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
 			    unsigned int length)
 {
+	struct rxe_wq *sq = &qp->sq;
 	enum ibv_wr_opcode opcode = ibwr->opcode;
 
 	if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
 	if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
 		return -EINVAL;
 
+	if (ibwr->opcode == IBV_WR_BIND_MW) {
+		if (length)
+			return -EINVAL;
+		if (ibwr->num_sge)
+			return -EINVAL;
+		if (ibwr->imm_data)
+			return -EINVAL;
+		if ((qp_type(qp) != IBV_QPT_RC) &&
+		    (qp_type(qp) != IBV_QPT_UC))
+			return -EINVAL;
+	}
+
 	return 0;
 }
 
 static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 {
+	struct ibv_mw *ibmw;
+	struct ibv_mr *ibmr;
+
 	memset(kwr, 0, sizeof(*kwr));
 
 	kwr->wr_id		= uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 		kwr->wr.atomic.rkey		= uwr->wr.atomic.rkey;
 		break;
 
+	case IBV_WR_BIND_MW:
+		ibmr = uwr->bind_mw.bind_info.mr;
+		ibmw = uwr->bind_mw.mw;
+
+		kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+		kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+		kwr->wr.mw.mr_lkey = ibmr->lkey;
+		kwr->wr.mw.mw_rkey = ibmw->rkey;
+		kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+		kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+		break;
+
 	default:
 		break;
 	}
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
 	if (ibwr->send_flags & IBV_SEND_INLINE) {
 		uint8_t *inline_data = wqe->dma.inline_data;
 
+		wqe->dma.resid = 0;
+
 		for (i = 0; i < num_sge; i++) {
 			memcpy(inline_data,
 			       (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
 		wqe->iova	= ibwr->wr.atomic.remote_addr;
 	else
 		wqe->iova	= ibwr->wr.rdma.remote_addr;
+
 	wqe->dma.length		= length;
 	wqe->dma.resid		= length;
 	wqe->dma.num_sge	= num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
 	for (i = 0; i < ibwr->num_sge; i++)
 		length += ibwr->sg_list[i].length;
 
-	err = validate_send_wr(sq, ibwr, length);
+	err = validate_send_wr(qp, ibwr, length);
 	if (err) {
 		printf("validate send failed\n");
 		return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
 	.dealloc_pd = rxe_dealloc_pd,
 	.reg_mr = rxe_reg_mr,
 	.dereg_mr = rxe_dereg_mr,
+	.alloc_mw = rxe_alloc_mw,
+	.dealloc_mw = rxe_dealloc_mw,
+	.bind_mw = rxe_bind_mw,
 	.create_cq = rxe_create_cq,
 	.create_cq_ex = rxe_create_cq_ex,
 	.poll_cq = rxe_poll_cq,
Signature · 1 line
2.30.2
#3 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 11 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
    u.poll_cq(server.cq)
  File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
    raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Quoted 259 lines
---
 providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
        return ret;
 }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
 static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                 uint64_t hca_va, int access)
 {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
        advance_qp_cur_index(qp);
 }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
 static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
 {
        struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                | IBV_QP_EX_WITH_SEND
                | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
        if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
        if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
 }
 
 /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                            unsigned int length)
 {
+       struct rxe_wq *sq = &qp->sq;
        enum ibv_wr_opcode opcode = ibwr->opcode;
 
        if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
        if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
        return 0;
 }
 
 static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
        memset(kwr, 0, sizeof(*kwr));
 
        kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
        default:
                break;
        }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
        if (ibwr->send_flags & IBV_SEND_INLINE) {
                uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                for (i = 0; i < num_sge; i++) {
                        memcpy(inline_data,
                               (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                wqe->iova       = ibwr->wr.atomic.remote_addr;
        else
                wqe->iova       = ibwr->wr.rdma.remote_addr;
+
        wqe->dma.length         = length;
        wqe->dma.resid          = length;
        wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
        for (i = 0; i < ibwr->num_sge; i++)
                length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
        if (err) {
                printf("validate send failed\n");
                return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
        .dealloc_pd = rxe_dealloc_pd,
        .reg_mr = rxe_reg_mr,
        .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
        .create_cq = rxe_create_cq,
        .create_cq_ex = rxe_create_cq_ex,
        .poll_cq = rxe_poll_cq,
--

2.30.2

#4 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 34 lines · Zhu Yanjun
Quoted 11 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
    u.poll_cq(server.cq)
  File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
    raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address
Quoted 6 lines

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Yes. It is because the test is not setting the BIND_MW access for the MR. It is not permitted to bind an MW to a MR that does not have the BIND_MW access flag set. But this test mistakenly thinks it is going to work. The other MW tests in test_mr.py do set the BIND_MW access flag. Bob

Quoted 260 lines
Quoted 260 lines
---
 providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
        return ret;
 }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
 static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                 uint64_t hca_va, int access)
 {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
        advance_qp_cur_index(qp);
 }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
 static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
 {
        struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                | IBV_QP_EX_WITH_SEND
                | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
        if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
        if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
 }
 
 /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                            unsigned int length)
 {
+       struct rxe_wq *sq = &qp->sq;
        enum ibv_wr_opcode opcode = ibwr->opcode;
 
        if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
        if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
        return 0;
 }
 
 static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
        memset(kwr, 0, sizeof(*kwr));
 
        kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
        default:
                break;
        }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
        if (ibwr->send_flags & IBV_SEND_INLINE) {
                uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                for (i = 0; i < num_sge; i++) {
                        memcpy(inline_data,
                               (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                wqe->iova       = ibwr->wr.atomic.remote_addr;
        else
                wqe->iova       = ibwr->wr.rdma.remote_addr;
+
        wqe->dma.length         = length;
        wqe->dma.resid          = length;
        wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
        for (i = 0; i < ibwr->num_sge; i++)
                length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
        if (err) {
                printf("validate send failed\n");
                return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
        .dealloc_pd = rxe_dealloc_pd,
        .reg_mr = rxe_reg_mr,
        .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
        .create_cq = rxe_create_cq,
        .create_cq_ex = rxe_create_cq_ex,
        .poll_cq = rxe_poll_cq,
--

2.30.2

#5 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 36 lines · Bob Pearson
Quoted 34 lines · Zhu Yanjun
Quoted 11 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
    u.poll_cq(server.cq)
  File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
    raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address
Quoted 7 lines
Quoted 6 lines

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Yes. It is because the test is not setting the BIND_MW access for the MR. It is not permitted to bind an MW to a MR that does not have the BIND_MW access flag set. But this test mistakenly thinks it is going to work. The other MW tests in test_mr.py do set the BIND_MW access flag.

Thanks a lot.
If the root cause to this problem is correct, and setting the BIND_MW access for the MR can fix this problem,

Please file a commit to fix this problem.

I am fine with this MW patch series if all the problems (including rdma-core and rxe in kernel) are fixed.

Thanks for your effort.

Zhu Yanjun

Quoted 262 lines

Bob

Quoted 260 lines
Quoted 260 lines
---
 providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
        return ret;
 }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
 static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                 uint64_t hca_va, int access)
 {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
        advance_qp_cur_index(qp);
 }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
 static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
 {
        struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                | IBV_QP_EX_WITH_SEND
                | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                | IBV_QP_EX_WITH_SEND_WITH_INV,
 
        RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
        if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
        if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
 }
 
 /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                            unsigned int length)
 {
+       struct rxe_wq *sq = &qp->sq;
        enum ibv_wr_opcode opcode = ibwr->opcode;
 
        if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
        if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
        return 0;
 }
 
 static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
        memset(kwr, 0, sizeof(*kwr));
 
        kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
        default:
                break;
        }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
        if (ibwr->send_flags & IBV_SEND_INLINE) {
                uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                for (i = 0; i < num_sge; i++) {
                        memcpy(inline_data,
                               (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                wqe->iova       = ibwr->wr.atomic.remote_addr;
        else
                wqe->iova       = ibwr->wr.rdma.remote_addr;
+
        wqe->dma.length         = length;
        wqe->dma.resid          = length;
        wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
        for (i = 0; i < ibwr->num_sge; i++)
                length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
        if (err) {
                printf("validate send failed\n");
                return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
        .dealloc_pd = rxe_dealloc_pd,
        .reg_mr = rxe_reg_mr,
        .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
        .create_cq = rxe_create_cq,
        .create_cq_ex = rxe_create_cq_ex,
        .poll_cq = rxe_poll_cq,
--

2.30.2

#6 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 45 lines · Zhu Yanjun
Quoted 33 lines · Bob Pearson
Quoted 32 lines · Zhu Yanjun
Quoted 10 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
   File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
     u.poll_cq(server.cq)
   File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
     raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address
Quoted 7 lines
Quoted 6 lines

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Yes. It is because the test is not setting the BIND_MW access for the MR. It is not permitted to bind an MW to a MR that does not have the BIND_MW access flag set. But this test mistakenly thinks it is going to work. The other MW tests in test_mr.py do set the BIND_MW access flag.

Thanks a lot.
If the root cause to this problem is correct, and setting the BIND_MW access for the MR can fix this problem,

Quoted 8 lines

Please file a commit to fix this problem.

I am fine with this MW patch series if all the problems (including rdma-core and rxe in kernel) are fixed.

Thanks for your effort.

Zhu Yanjun

Zhu,

There was a second test bug, now fixed. And all the python tests are passing or skipped. No errors. There is a message to the liux-rdma list from Edward Srouji indicating that the patch is upstream at github or you can use the one in my email to him. It will be a while until his PR goes up stream I would guess.

I sent a fresh set of patches last night for both the kernel and user space provider.

Bob

Quoted 260 lines
Quoted 260 lines

Bob

Quoted 259 lines
Quoted 259 lines
---
  providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
  1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
         return ret;
  }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
  static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                  uint64_t hca_va, int access)
  {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
         advance_qp_cur_index(qp);
  }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
  static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
  {
         struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                 | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                 | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                 | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                 | IBV_QP_EX_WITH_SEND_WITH_INV,
 
         RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                 | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                 | IBV_QP_EX_WITH_SEND
                 | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                 | IBV_QP_EX_WITH_SEND_WITH_INV,
 
         RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
         if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                 qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
         if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                 qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
  }
 
  /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                             unsigned int length)
  {
+       struct rxe_wq *sq = &qp->sq;
         enum ibv_wr_opcode opcode = ibwr->opcode;
 
         if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
         if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                 return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
         return 0;
  }
 
  static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
  {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
         memset(kwr, 0, sizeof(*kwr));
 
         kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                 kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                 break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
         default:
                 break;
         }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
         if (ibwr->send_flags & IBV_SEND_INLINE) {
                 uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                 for (i = 0; i < num_sge; i++) {
                         memcpy(inline_data,
                                (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                 wqe->iova       = ibwr->wr.atomic.remote_addr;
         else
                 wqe->iova       = ibwr->wr.rdma.remote_addr;
+
         wqe->dma.length         = length;
         wqe->dma.resid          = length;
         wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
         for (i = 0; i < ibwr->num_sge; i++)
                 length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
         if (err) {
                 printf("validate send failed\n");
                 return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
         .dealloc_pd = rxe_dealloc_pd,
         .reg_mr = rxe_reg_mr,
         .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
         .create_cq = rxe_create_cq,
         .create_cq_ex = rxe_create_cq_ex,
         .poll_cq = rxe_poll_cq,
--

2.30.2

#7 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 59 lines · Robert B
Quoted 45 lines · Zhu Yanjun
Quoted 33 lines · Bob Pearson
Quoted 32 lines · Zhu Yanjun
Quoted 10 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
   File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
     u.poll_cq(server.cq)
   File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
     raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address
Quoted 7 lines
Quoted 6 lines

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Yes. It is because the test is not setting the BIND_MW access for the MR. It is not permitted to bind an MW to a MR that does not have the BIND_MW access flag set. But this test mistakenly thinks it is going to work. The other MW tests in test_mr.py do set the BIND_MW access flag.

Thanks a lot.
If the root cause to this problem is correct, and setting the BIND_MW access for the MR can fix this problem,

Quoted 8 lines

Please file a commit to fix this problem.

I am fine with this MW patch series if all the problems (including rdma-core and rxe in kernel) are fixed.

Thanks for your effort.

Zhu Yanjun

Zhu,

Quoted 8 lines

There was a second test bug, now fixed. And all the python tests are passing or skipped. No errors. There is a message to the liux-rdma list from Edward Srouji indicating that the patch is upstream at github or you can use the one in my email to him. It will be a while until his PR goes up stream I would guess.

I sent a fresh set of patches last night for both the kernel and user space provider.

Thanks a lot for your effort.

I applied your commits for rdma-core and rxe in kernel, and the following diff.

All the python tests in rdma-core pass or skipped. No errors

Thanks again.

"

tests/test_qpex.py +1 -1
index 20288d45..41028e2d 100644
@@ -146,10 +146,10 @@ class QpExRCAtomicFetchAdd(RCResources):
 
 class QpExRCBindMw(RCResources):
     def create_qps(self):
-        create_qp_ex(self, e.IBV_QPT_RC, e.IBV_QP_EX_WITH_BIND_MW)
+        create_qp_ex(self, e.IBV_QPT_RC,

e.IBV_QP_EX_WITH_BIND_MW|e.IBV_QP_EX_WITH_RDMA_WRITE)

def create_mr(self):
- self.mr = u.create_custom_mr(self, e.IBV_ACCESS_REMOTE_WRITE)
+ self.mr = u.create_custom_mr(self,
e.IBV_ACCESS_REMOTE_WRITE|e.IBV_ACCESS_MW_BIND)
"

Zhu Yanjun

Quoted 263 lines

Bob

Quoted 260 lines
Quoted 260 lines

Bob

Quoted 259 lines
Quoted 259 lines
---
  providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
  1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
         return ret;
  }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
  static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                  uint64_t hca_va, int access)
  {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
         advance_qp_cur_index(qp);
  }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
  static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
  {
         struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                 | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                 | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                 | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                 | IBV_QP_EX_WITH_SEND_WITH_INV,
 
         RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                 | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                 | IBV_QP_EX_WITH_SEND
                 | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                 | IBV_QP_EX_WITH_SEND_WITH_INV,
 
         RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
         if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                 qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
         if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                 qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
  }
 
  /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                             unsigned int length)
  {
+       struct rxe_wq *sq = &qp->sq;
         enum ibv_wr_opcode opcode = ibwr->opcode;
 
         if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
         if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                 return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
         return 0;
  }
 
  static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
  {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
         memset(kwr, 0, sizeof(*kwr));
 
         kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                 kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                 break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
         default:
                 break;
         }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
         if (ibwr->send_flags & IBV_SEND_INLINE) {
                 uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                 for (i = 0; i < num_sge; i++) {
                         memcpy(inline_data,
                                (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                 wqe->iova       = ibwr->wr.atomic.remote_addr;
         else
                 wqe->iova       = ibwr->wr.rdma.remote_addr;
+
         wqe->dma.length         = length;
         wqe->dma.resid          = length;
         wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
         for (i = 0; i < ibwr->num_sge; i++)
                 length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
         if (err) {
                 printf("validate send failed\n");
                 return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
         .dealloc_pd = rxe_dealloc_pd,
         .reg_mr = rxe_reg_mr,
         .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
         .create_cq = rxe_create_cq,
         .create_cq_ex = rxe_create_cq_ex,
         .poll_cq = rxe_poll_cq,
--

2.30.2

#8 lore raw

Re: [PATCH v2 2/2] Providers/rxe: Implement memory windows

Quoted 66 lines · Zhu Yanjun
Quoted 57 lines · Robert B
Quoted 45 lines · Zhu Yanjun
Quoted 33 lines · Bob Pearson
Quoted 32 lines · Zhu Yanjun
Quoted 10 lines · Bob Pearson

This patch makes the required changes to the rxe provider to support the kernel memory windows patches to the rxe driver.

The following changes are made:
- Add ibv_alloc_mw verb
- Add ibv_dealloc_mw verb
- Add ibv_bind_mw verb for type 1 MWs
- Add support for bind MW send work requests through the traditional QP API and the extended QP API.

Quoted 4 lines

---
v2:
Added support for extended QP bind MW work requests.

I got the following errors.
Is it a known issue?

# ./bin/run_tests.py --dev rxe0
.............sssssssss.............ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss........sssssssssssssssssss....ssss.........s...s.....E.......ssssssssss..ss
======================================================================
ERROR: test_qp_ex_rc_bind_mw (tests.test_qpex.QpExTestCase)
Verify bind memory window operation using the new post_send API.
----------------------------------------------------------------------
Traceback (most recent call last):
    File "/root/rdma-core/tests/test_qpex.py", line 292, in test_qp_ex_rc_bind_mw
      u.poll_cq(server.cq)
    File "/root/rdma-core/tests/utils.py", line 538, in poll_cq
      raise PyverbsRDMAError('Completion status is {s}'.
pyverbs.pyverbs_error.PyverbsRDMAError: Completion status is Memory
window bind error. Errno: 6, No such device or address
Quoted 7 lines
Quoted 6 lines

----------------------------------------------------------------------
Ran 193 tests in 2.544s

FAILED (errors=1, skipped=128)

Zhu Yanjun

Yes. It is because the test is not setting the BIND_MW access for the MR. It is not permitted to bind an MW to a MR that does not have the BIND_MW access flag set. But this test mistakenly thinks it is going to work. The other MW tests in test_mr.py do set the BIND_MW access flag.

Thanks a lot.
If the root cause to this problem is correct, and setting the BIND_MW access for the MR can fix this problem,

Quoted 8 lines

Please file a commit to fix this problem.

I am fine with this MW patch series if all the problems (including rdma-core and rxe in kernel) are fixed.

Thanks for your effort.

Zhu Yanjun

Zhu,

Quoted 8 lines

There was a second test bug, now fixed. And all the python tests are passing or skipped. No errors. There is a message to the liux-rdma list from Edward Srouji indicating that the patch is upstream at github or you can use the one in my email to him. It will be a while until his PR goes up stream I would guess.

I sent a fresh set of patches last night for both the kernel and user space provider.

Quoted 7 lines

Thanks a lot for your effort.

I applied your commits for rdma-core and rxe in kernel, and the following diff.

All the python tests in rdma-core pass or skipped. No errors

Thanks again.

Thanks to you for your patience.

Bob

Quoted 284 lines

"

tests/test_qpex.py +2 -2
index 20288d45..41028e2d 100644
@@ -146,10 +146,10 @@ class QpExRCAtomicFetchAdd(RCResources):
 
  class QpExRCBindMw(RCResources):
      def create_qps(self):
-        create_qp_ex(self, e.IBV_QPT_RC, e.IBV_QP_EX_WITH_BIND_MW)
+        create_qp_ex(self, e.IBV_QPT_RC,
 e.IBV_QP_EX_WITH_BIND_MW|e.IBV_QP_EX_WITH_RDMA_WRITE)
 
      def create_mr(self):
-        self.mr = u.create_custom_mr(self, e.IBV_ACCESS_REMOTE_WRITE)
+        self.mr = u.create_custom_mr(self,
 e.IBV_ACCESS_REMOTE_WRITE|e.IBV_ACCESS_MW_BIND)

"

Zhu Yanjun

Quoted 262 lines

Bob

Quoted 260 lines
Quoted 260 lines

Bob

Quoted 259 lines
Quoted 259 lines
---
   providers/rxe/rxe.c | 157 +++++++++++++++++++++++++++++++++++++++++++-
   1 file changed, 155 insertions(+), 2 deletions(-)
providers/rxe/rxe.c +155 -3
index a68656ae..bb39ef04 100644
@@ -128,6 +128,95 @@ static int rxe_dealloc_pd(struct ibv_pd *pd)
          return ret;
   }
 
+static struct ibv_mw *rxe_alloc_mw(struct ibv_pd *ibpd, enum ibv_mw_type type)
+{
+       int ret;
+       struct ibv_mw *ibmw;
+       struct ibv_alloc_mw cmd = {};
+       struct ib_uverbs_alloc_mw_resp resp = {};
+
+       ibmw = calloc(1, sizeof(*ibmw));
+       if (!ibmw)
+               return NULL;
+
+       ret = ibv_cmd_alloc_mw(ibpd, type, ibmw, &cmd, sizeof(cmd),
+                                               &resp, sizeof(resp));
+       if (ret) {
+               free(ibmw);
+               return NULL;
+       }
+
+       return ibmw;
+}
+
+static int rxe_dealloc_mw(struct ibv_mw *ibmw)
+{
+       int ret;
+
+       ret = ibv_cmd_dealloc_mw(ibmw);
+       if (ret)
+               return ret;
+
+       free(ibmw);
+       return 0;
+}
+
+static int next_rkey(int rkey)
+{
+       return (rkey & 0xffffff00) | ((rkey + 1) & 0x000000ff);
+}
+
+static int rxe_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr_list,
+                        struct ibv_send_wr **bad_wr);
+
+static int rxe_bind_mw(struct ibv_qp *ibqp, struct ibv_mw *ibmw,
+                       struct ibv_mw_bind *mw_bind)
+{
+       int ret;
+       struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info;
+       struct ibv_send_wr ibwr;
+       struct ibv_send_wr *bad_wr;
+
+       if (!bind_info->mr && (bind_info->addr || bind_info->length)) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mw_access_flags & IBV_ACCESS_ZERO_BASED) {
+               ret = EINVAL;
+               goto err;
+       }
+
+       if (bind_info->mr) {
+               if (ibmw->pd != bind_info->mr->pd) {
+                       ret = EPERM;
+                       goto err;
+               }
+       }
+
+       memset(&ibwr, 0, sizeof(ibwr));
+
+       ibwr.opcode             = IBV_WR_BIND_MW;
+       ibwr.next               = NULL;
+       ibwr.wr_id              = mw_bind->wr_id;
+       ibwr.send_flags         = mw_bind->send_flags;
+       ibwr.bind_mw.bind_info  = mw_bind->bind_info;
+       ibwr.bind_mw.mw         = ibmw;
+       ibwr.bind_mw.rkey       = next_rkey(ibmw->rkey);
+
+       ret = rxe_post_send(ibqp, &ibwr, &bad_wr);
+       if (ret)
+               goto err;
+
+       /* user has to undo this if he gets an error wc */
+       ibmw->rkey = ibwr.bind_mw.rkey;
+
+       return 0;
+err:
+       errno = ret;
+       return errno;
+}
+
   static struct ibv_mr *rxe_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
                                   uint64_t hca_va, int access)
   {
@@ -715,6 +804,31 @@ static void wr_atomic_fetch_add(struct ibv_qp_ex *ibqp, uint32_t rkey,
          advance_qp_cur_index(qp);
   }
 
+static void wr_bind_mw(struct ibv_qp_ex *ibqp, struct ibv_mw *ibmw,
+                      uint32_t rkey, const struct ibv_mw_bind_info *info)
+{
+       struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
+       struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue, qp->cur_index);
+
+       if (check_qp_queue_full(qp))
+               return;
+
+       memset(wqe, 0, sizeof(*wqe));
+
+       wqe->wr.wr_id = ibqp->wr_id;
+       wqe->wr.opcode = IBV_WR_BIND_MW;
+       wqe->wr.send_flags = qp->vqp.qp_ex.wr_flags;
+       wqe->wr.wr.mw.addr = info->addr;
+       wqe->wr.wr.mw.length = info->length;
+       wqe->wr.wr.mw.mr_lkey = info->mr->lkey;
+       wqe->wr.wr.mw.mw_rkey = ibmw->rkey;
+       wqe->wr.wr.mw.rkey = rkey;
+       wqe->wr.wr.mw.access = info->mw_access_flags;
+       wqe->ssn = qp->ssn++;
+
+       advance_qp_cur_index(qp);
+}
+
   static void wr_local_inv(struct ibv_qp_ex *ibqp, uint32_t invalidate_rkey)
   {
          struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
@@ -1106,6 +1220,7 @@ enum {
                  | IBV_QP_EX_WITH_ATOMIC_CMP_AND_SWP
                  | IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD
                  | IBV_QP_EX_WITH_LOCAL_INV
+               | IBV_QP_EX_WITH_BIND_MW
                  | IBV_QP_EX_WITH_SEND_WITH_INV,
 
          RXE_SUP_UC_QP_SEND_OPS_FLAGS =
@@ -1113,6 +1228,7 @@ enum {
                  | IBV_QP_EX_WITH_RDMA_WRITE_WITH_IMM
                  | IBV_QP_EX_WITH_SEND
                  | IBV_QP_EX_WITH_SEND_WITH_IMM
+               | IBV_QP_EX_WITH_BIND_MW
                  | IBV_QP_EX_WITH_SEND_WITH_INV,
 
          RXE_SUP_UD_QP_SEND_OPS_FLAGS =
@@ -1162,6 +1278,9 @@ static void set_qp_send_ops(struct rxe_qp *qp, uint64_t flags)
          if (flags & IBV_QP_EX_WITH_ATOMIC_FETCH_AND_ADD)
                  qp->vqp.qp_ex.wr_atomic_fetch_add = wr_atomic_fetch_add;
 
+       if (flags & IBV_QP_EX_WITH_BIND_MW)
+               qp->vqp.qp_ex.wr_bind_mw = wr_bind_mw;
+
          if (flags & IBV_QP_EX_WITH_LOCAL_INV)
                  qp->vqp.qp_ex.wr_local_inv = wr_local_inv;
 
@@ -1275,9 +1394,10 @@ static int rxe_destroy_qp(struct ibv_qp *ibqp)
   }
 
   /* basic sanity checks for send work request */
-static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
+static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
                              unsigned int length)
   {
+       struct rxe_wq *sq = &qp->sq;
          enum ibv_wr_opcode opcode = ibwr->opcode;
 
          if (ibwr->num_sge > sq->max_sge)
@@ -1291,11 +1411,26 @@ static int validate_send_wr(struct rxe_wq *sq, struct ibv_send_wr *ibwr,
          if ((ibwr->send_flags & IBV_SEND_INLINE) && (length > sq->max_inline))
                  return -EINVAL;
 
+       if (ibwr->opcode == IBV_WR_BIND_MW) {
+               if (length)
+                       return -EINVAL;
+               if (ibwr->num_sge)
+                       return -EINVAL;
+               if (ibwr->imm_data)
+                       return -EINVAL;
+               if ((qp_type(qp) != IBV_QPT_RC) &&
+                   (qp_type(qp) != IBV_QPT_UC))
+                       return -EINVAL;
+       }
+
          return 0;
   }
 
   static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
   {
+       struct ibv_mw *ibmw;
+       struct ibv_mr *ibmr;
+
          memset(kwr, 0, sizeof(*kwr));
 
          kwr->wr_id              = uwr->wr_id;
@@ -1326,6 +1461,18 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
                  kwr->wr.atomic.rkey             = uwr->wr.atomic.rkey;
                  break;
 
+       case IBV_WR_BIND_MW:
+               ibmr = uwr->bind_mw.bind_info.mr;
+               ibmw = uwr->bind_mw.mw;
+
+               kwr->wr.mw.addr = uwr->bind_mw.bind_info.addr;
+               kwr->wr.mw.length = uwr->bind_mw.bind_info.length;
+               kwr->wr.mw.mr_lkey = ibmr->lkey;
+               kwr->wr.mw.mw_rkey = ibmw->rkey;
+               kwr->wr.mw.rkey = uwr->bind_mw.rkey;
+               kwr->wr.mw.access = uwr->bind_mw.bind_info.mw_access_flags;
+               break;
+
          default:
                  break;
          }
@@ -1348,6 +1495,8 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
          if (ibwr->send_flags & IBV_SEND_INLINE) {
                  uint8_t *inline_data = wqe->dma.inline_data;
 
+               wqe->dma.resid = 0;
+
                  for (i = 0; i < num_sge; i++) {
                          memcpy(inline_data,
                                 (uint8_t *)(long)ibwr->sg_list[i].addr,
@@ -1363,6 +1512,7 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
                  wqe->iova       = ibwr->wr.atomic.remote_addr;
          else
                  wqe->iova       = ibwr->wr.rdma.remote_addr;
+
          wqe->dma.length         = length;
          wqe->dma.resid          = length;
          wqe->dma.num_sge        = num_sge;
@@ -1385,7 +1535,7 @@ static int post_one_send(struct rxe_qp *qp, struct rxe_wq *sq,
          for (i = 0; i < ibwr->num_sge; i++)
                  length += ibwr->sg_list[i].length;
 
-       err = validate_send_wr(sq, ibwr, length);
+       err = validate_send_wr(qp, ibwr, length);
          if (err) {
                  printf("validate send failed\n");
                  return err;
@@ -1579,6 +1729,9 @@ static const struct verbs_context_ops rxe_ctx_ops = {
          .dealloc_pd = rxe_dealloc_pd,
          .reg_mr = rxe_reg_mr,
          .dereg_mr = rxe_dereg_mr,
+       .alloc_mw = rxe_alloc_mw,
+       .dealloc_mw = rxe_dealloc_mw,
+       .bind_mw = rxe_bind_mw,
          .create_cq = rxe_create_cq,
          .create_cq_ex = rxe_create_cq_ex,
          .poll_cq = rxe_poll_cq,
--

2.30.2