prov/efa: Remove inline write logic for rma inject

We previously set the inject size to 0 to prevent using inline write,
but fabtests can use inject rma for 0 size message.
Make rma inject temporarily return FI_ENOSYS before firmware
supports inline write.

Signed-off-by: Jessie Yang <[email protected]>

Author: jiaxiyan

prov/efa/src/efa_rma.c

@@ -200,17 +200,21 @@ static inline ssize_t efa_rma_post_write(struct efa_base_ep *base_ep,
 	struct efa_conn *conn;
 #ifndef _WIN32
 	struct ibv_sge sge_list[msg->iov_count];
-	struct ibv_data_buf inline_data_list[msg->iov_count];
 #else
 	/* MSVC compiler does not support array declarations with runtime size, so hardcode
 	 * the expected iov_limit/max_sq_sge from the lower-level efa provider.
 	 */
 	struct ibv_sge sge_list[EFA_DEV_ATTR_MAX_WR_SGE];
 	struct ibv_data_buf inline_data_list[EFA_DEV_ATTR_MAX_WR_SGE];
 #endif
-	size_t len;
 	int i, err = 0;
 
+	if (flags & FI_INJECT) {
+		EFA_WARN(FI_LOG_EP_DATA,
+			 "FI_INJECT is not supported by efa rma yet.\n");
+		return -FI_ENOSYS;
+	}
+
 	efa_tracepoint(write_begin_msg_context, (size_t) msg->context, (size_t) msg->addr);
 
 	qp = base_ep->qp;
@@ -230,24 +234,13 @@ static inline ssize_t efa_rma_post_write(struct efa_base_ep *base_ep,
 		ibv_wr_rdma_write(qp->ibv_qp_ex, msg->rma_iov[0].key, msg->rma_iov[0].addr);
 	}
 
-	len = ofi_total_iov_len(msg->msg_iov, msg->iov_count);
-	if (len <= base_ep->domain->device->efa_attr.inline_buf_size &&
-	    len <= base_ep->inject_rma_size &&
-	    (!msg->desc || !efa_mr_is_hmem(msg->desc[0]))) {
-		for (i = 0; i < msg->iov_count; i++) {
-			inline_data_list[i].addr = msg->msg_iov[i].iov_base;
-			inline_data_list[i].length = msg->msg_iov[i].iov_len;
-		}
-		ibv_wr_set_inline_data_list(qp->ibv_qp_ex, msg->iov_count, inline_data_list);
-	} else {
-		for (i = 0; i < msg->iov_count; ++i) {
-			sge_list[i].addr = (uint64_t)msg->msg_iov[i].iov_base;
-			sge_list[i].length = msg->msg_iov[i].iov_len;
-			assert(msg->desc && msg->desc[i]);
-			sge_list[i].lkey = ((struct efa_mr *)msg->desc[i])->ibv_mr->lkey;
-		}
-		ibv_wr_set_sge_list(qp->ibv_qp_ex, msg->iov_count, sge_list);
+	for (i = 0; i < msg->iov_count; ++i) {
+		sge_list[i].addr = (uint64_t)msg->msg_iov[i].iov_base;
+		sge_list[i].length = msg->msg_iov[i].iov_len;
+		assert(msg->desc && msg->desc[i]);
+		sge_list[i].lkey = ((struct efa_mr *)msg->desc[i])->ibv_mr->lkey;
 	}
+	ibv_wr_set_sge_list(qp->ibv_qp_ex, msg->iov_count, sge_list);
 
 	conn = efa_av_addr_to_conn(base_ep->av, msg->addr);
 	assert(conn && conn->ep_addr);
@@ -348,51 +341,6 @@ ssize_t efa_rma_writedata(struct fid_ep *ep_fid, const void *buf, size_t len,
 	return efa_rma_post_write(base_ep, &msg, FI_REMOTE_CQ_DATA | efa_tx_flags(base_ep));
 }
 
-ssize_t efa_rma_inject_write(struct fid_ep *ep_fid, const void *buf, size_t len,
-			     fi_addr_t dest_addr, uint64_t addr, uint64_t key)
-{
-	struct fi_msg_rma msg;
-	struct iovec iov;
-	struct fi_rma_iov rma_iov;
-	struct efa_base_ep *base_ep;
-	int err;
-
-	base_ep = container_of(ep_fid, struct efa_base_ep, util_ep.ep_fid);
-	assert(len <= base_ep->inject_rma_size);
-	err = efa_rma_check_cap(base_ep);
-	if (err)
-		return err;
-
-	EFA_SETUP_IOV(iov, buf, len);
-	EFA_SETUP_RMA_IOV(rma_iov, addr, len, key);
-	EFA_SETUP_MSG_RMA(msg, &iov, NULL, 1, dest_addr, &rma_iov, 1, NULL, 0);
-
-	return efa_rma_post_write(base_ep, &msg, FI_INJECT);
-}
-
-ssize_t efa_rma_inject_writedata(struct fid_ep *ep_fid, const void *buf,
-				 size_t len, uint64_t data, fi_addr_t dest_addr,
-				 uint64_t addr, uint64_t key)
-{
-	struct fi_msg_rma msg;
-	struct iovec iov;
-	struct fi_rma_iov rma_iov;
-	struct efa_base_ep *base_ep;
-	int err;
-
-	base_ep = container_of(ep_fid, struct efa_base_ep, util_ep.ep_fid);
-	assert(len <= base_ep->inject_rma_size);
-	err = efa_rma_check_cap(base_ep);
-	if (err)
-		return err;
-
-	EFA_SETUP_IOV(iov, buf, len);
-	EFA_SETUP_RMA_IOV(rma_iov, addr, len, key);
-	EFA_SETUP_MSG_RMA(msg, &iov, NULL, 1, dest_addr, &rma_iov, 1, NULL, data);
-
-	return efa_rma_post_write(base_ep, &msg, FI_INJECT | FI_REMOTE_CQ_DATA);
-}
-
 struct fi_ops_rma efa_dgram_ep_rma_ops = {
 	.size = sizeof(struct fi_ops_rma),
 	.read = fi_no_rma_read,
@@ -414,7 +362,7 @@ struct fi_ops_rma efa_rma_ops = {
 	.write = efa_rma_write,
 	.writev = efa_rma_writev,
 	.writemsg = efa_rma_writemsg,
-	.inject = efa_rma_inject_write,
+	.inject = fi_no_rma_inject,
 	.writedata = efa_rma_writedata,
-	.injectdata = efa_rma_inject_writedata,
+	.injectdata = fi_no_rma_injectdata,
 };

prov/efa/test/efa_unit_test_rma.c

@@ -25,8 +25,6 @@ static void test_efa_rma_prep(struct efa_resource *resource, fi_addr_t *addr)
 	ibv_qpx->wr_rdma_read = &efa_mock_ibv_wr_rdma_read_save_wr;
 	ibv_qpx->wr_rdma_write = &efa_mock_ibv_wr_rdma_write_save_wr;
 	ibv_qpx->wr_rdma_write_imm = &efa_mock_ibv_wr_rdma_write_imm_save_wr;
-	ibv_qpx->wr_set_inline_data_list =
-		&efa_mock_ibv_wr_set_inline_data_list_no_op;
 	ibv_qpx->wr_set_sge_list = &efa_mock_ibv_wr_set_sge_list_no_op;
 	ibv_qpx->wr_set_ud_addr = &efa_mock_ibv_wr_set_ud_addr_no_op;
 	ibv_qpx->wr_complete = &efa_mock_ibv_wr_complete_no_op;
@@ -241,11 +239,9 @@ void test_efa_rma_inject_write(struct efa_resource **state)
 	test_efa_rma_prep(resource, &dest_addr);
 	efa_unit_test_buff_construct(&local_buff, resource, 32 /* buff_size */);
 
-	assert_int_equal(g_ibv_submitted_wr_id_cnt, 0);
 	ret = fi_inject_write(resource->ep, local_buff.buff, local_buff.size,
 			      dest_addr, remote_addr, remote_key);
-	assert_int_equal(ret, 0);
-	assert_int_equal(g_ibv_submitted_wr_id_cnt, 1);
+	assert_int_equal(ret, -FI_ENOSYS);
 
 	efa_unit_test_buff_destruct(&local_buff);
 }
@@ -262,12 +258,39 @@ void test_efa_rma_inject_writedata(struct efa_resource **state)
 	test_efa_rma_prep(resource, &dest_addr);
 	efa_unit_test_buff_construct(&local_buff, resource, 32 /* buff_size */);
 
-	assert_int_equal(g_ibv_submitted_wr_id_cnt, 0);
 	ret = fi_inject_writedata(resource->ep, local_buff.buff,
 				  local_buff.size, 0, dest_addr, remote_addr,
 				  remote_key);
-	assert_int_equal(ret, 0);
-	assert_int_equal(g_ibv_submitted_wr_id_cnt, 1);
+	assert_int_equal(ret, -FI_ENOSYS);
+
+	efa_unit_test_buff_destruct(&local_buff);
+}
+
+void test_efa_rma_writemsg_with_inject(struct efa_resource **state)
+{
+	struct efa_resource *resource = *state;
+	struct efa_unit_test_buff local_buff;
+	struct iovec iov;
+	struct fi_msg_rma msg = {0};
+	struct fi_rma_iov rma_iov;
+	fi_addr_t dest_addr;
+	void *desc;
+	int ret;
+
+	test_efa_rma_prep(resource, &dest_addr);
+	efa_unit_test_buff_construct(&local_buff, resource, 4096 /* buff_size */);
+
+	iov.iov_base = local_buff.buff;
+	iov.iov_len = local_buff.size;
+	desc = fi_mr_desc(local_buff.mr);
+	rma_iov.len = local_buff.size;
+	rma_iov.addr = 0x87654321;
+	rma_iov.key = 123456;
+	efa_unit_test_construct_msg_rma(&msg, &iov, &desc, 1, dest_addr, &rma_iov,
+					1, NULL, 0);
+
+	ret = fi_writemsg(resource->ep, &msg, FI_INJECT);
+	assert_int_equal(ret, -FI_ENOSYS);
 
 	efa_unit_test_buff_destruct(&local_buff);
 }

prov/efa/test/efa_unit_tests.c

@@ -229,6 +229,7 @@ int main(void)
 		cmocka_unit_test_setup_teardown(test_efa_rma_writedata, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rma_inject_write, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rma_inject_writedata, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
+		cmocka_unit_test_setup_teardown(test_efa_rma_writemsg_with_inject, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_cq_read_send_success, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_cq_read_recv_success, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_cq_read_send_failure, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),

prov/efa/test/efa_unit_tests.h

@@ -251,6 +251,7 @@ void test_efa_rma_writemsg();
 void test_efa_rma_writedata();
 void test_efa_rma_inject_write();
 void test_efa_rma_inject_writedata();
+void test_efa_rma_writemsg_with_inject();
 void test_efa_cq_read_send_success();
 void test_efa_cq_read_recv_success();
 void test_efa_cq_read_send_failure();