prov/efa: Require FI_MR_LOCAL for zero-copy receive

Zero-copy receive requires application to register the local buffer for transfer, otherwise libfabric will need to do a copy or registration internally, which doesn't comply with the zero-copy nature. Therefore, FI_MR_LOCAL is required for libfabric to turn on zcpy recv. Signed-off-by: Jessie Yang <[email protected]>
jiaxiyan · Aug 27, 2024 · 5e31d84 · 5e31d84
1 parent fb361d4
commit 5e31d84
Show file tree

Hide file tree

Showing 5 changed files with 43 additions and 4 deletions.
diff --git a/prov/efa/src/rdm/efa_rdm_ep_fiops.c b/prov/efa/src/rdm/efa_rdm_ep_fiops.c
@@ -468,6 +468,13 @@ void efa_rdm_ep_set_use_zcpy_rx(struct efa_rdm_ep *ep)
 		goto out;
 	}
 
+	/* FI_MR_LOCAL is not set, turn off zcpy recv */
+	if (!ep->base_ep.domain->mr_local) {
+		EFA_INFO(FI_LOG_EP_CTRL, "FI_MR_LOCAL mode bit is not set, zero-copy receive protocol will be disabled\n");
+		ep->use_zcpy_rx = false;
+		goto out;
+	}
+
 	if (ep->shm_ep) {
 		EFA_INFO(FI_LOG_EP_CTRL, "Libfabric SHM is not turned off, zero-copy receive protocol will be disabled\n");
 		ep->use_zcpy_rx = false;

diff --git a/prov/efa/src/rdm/efa_rdm_ep_utils.c b/prov/efa/src/rdm/efa_rdm_ep_utils.c
@@ -207,7 +207,7 @@ struct efa_rdm_ope *efa_rdm_ep_alloc_rxe(struct efa_rdm_ep *ep, fi_addr_t addr,
  */
 int efa_rdm_ep_post_user_recv_buf(struct efa_rdm_ep *ep, struct efa_rdm_ope *rxe, size_t flags)
 {
-	struct efa_rdm_pke *pkt_entry;
+	struct efa_rdm_pke *pkt_entry = NULL;
 	size_t rx_iov_offset = 0;
 	int err, rx_iov_index = 0;
 
@@ -225,29 +225,43 @@ int efa_rdm_ep_post_user_recv_buf(struct efa_rdm_ep *ep, struct efa_rdm_ope *rxe
 	err = ofi_iov_locate(rxe->iov, rxe->iov_count, ep->msg_prefix_size, &rx_iov_index, &rx_iov_offset);
 	if (OFI_UNLIKELY(err)) {
 		EFA_WARN(FI_LOG_CQ, "ofi_iov_locate failure: %s (%d)\n", fi_strerror(-err), -err);
-		return err;
+		goto err_free;
 	}
+
 	assert(rx_iov_index < rxe->iov_count);
 	assert(rx_iov_offset < rxe->iov[rx_iov_index].iov_len);
+	assert(ep->base_ep.domain->mr_local);
+
+	if (!rxe->desc[rx_iov_index]) {
+		err = -FI_EINVAL;
+		EFA_WARN(FI_LOG_EP_CTRL,
+			 "No valid desc is provided, not complied with FI_MR_LOCAL: %s (%d)\n",
+			 fi_strerror(-err), -err);
+		goto err_free;
+	}
 
 	pkt_entry->payload = (char *) rxe->iov[rx_iov_index].iov_base + rx_iov_offset;
 	pkt_entry->payload_mr = rxe->desc[rx_iov_index];
 	pkt_entry->payload_size = ofi_total_iov_len(&rxe->iov[rx_iov_index], rxe->iov_count - rx_iov_index) - rx_iov_offset;
 
 	err = efa_rdm_pke_recvv(&pkt_entry, 1);
 	if (OFI_UNLIKELY(err)) {
-		efa_rdm_pke_release_rx(pkt_entry);
 		EFA_WARN(FI_LOG_EP_CTRL,
 			"failed to post user supplied buffer %d (%s)\n", -err,
 			fi_strerror(-err));
-		return err;
+		goto err_free;
 	}
 
 #if ENABLE_DEBUG
 	dlist_insert_tail(&pkt_entry->dbg_entry, &ep->rx_posted_buf_list);
 #endif
 	ep->user_rx_pkts_posted++;
 	return 0;
+
+err_free:
+	if (pkt_entry)
+		efa_rdm_pke_release_rx(pkt_entry);
+	return err;
 }
 
 

diff --git a/prov/efa/test/efa_unit_test_ep.c b/prov/efa/test/efa_unit_test_ep.c
@@ -1066,6 +1066,22 @@ void test_efa_rdm_ep_user_p2p_not_supported_zcpy_rx_happy(struct efa_resource **
 	test_efa_rdm_ep_use_zcpy_rx_impl(resource, false, false, false);
 }
 
+/**
+ * @brief Verify zcpy_rx is disabled if FI_MR_LOCAL is not set
+ */
+void test_efa_rdm_ep_user_zcpy_rx_unhappy_due_to_no_mr_local(struct efa_resource **state)
+{
+	struct efa_resource *resource = *state;
+
+	resource->hints = efa_unit_test_alloc_hints(FI_EP_RDM);
+	assert_non_null(resource->hints);
+
+	resource->hints->caps = FI_MSG;
+	resource->hints->domain_attr->mr_mode &= ~FI_MR_LOCAL;
+
+	test_efa_rdm_ep_use_zcpy_rx_impl(resource, false, true, false);
+}
+
 void test_efa_rdm_ep_close_discard_posted_recv(struct efa_resource **state)
 {
 	struct efa_resource *resource = *state;

diff --git a/prov/efa/test/efa_unit_tests.c b/prov/efa/test/efa_unit_tests.c
@@ -108,6 +108,7 @@ int main(void)
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_user_disable_p2p_zcpy_rx_happy, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_user_zcpy_rx_unhappy_due_to_sas, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_user_p2p_not_supported_zcpy_rx_happy, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
+		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_user_zcpy_rx_unhappy_due_to_no_mr_local, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_close_discard_posted_recv, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_zcpy_recv_cancel, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_post_handshake_error_handling_pke_exhaustion, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),

diff --git a/prov/efa/test/efa_unit_tests.h b/prov/efa/test/efa_unit_tests.h
@@ -118,6 +118,7 @@ void test_efa_rdm_ep_user_zcpy_rx_disabled();
 void test_efa_rdm_ep_user_disable_p2p_zcpy_rx_happy();
 void test_efa_rdm_ep_user_zcpy_rx_unhappy_due_to_sas();
 void test_efa_rdm_ep_user_p2p_not_supported_zcpy_rx_happy();
+void test_efa_rdm_ep_user_zcpy_rx_unhappy_due_to_no_mr_local();
 void test_efa_rdm_ep_close_discard_posted_recv();
 void test_efa_rdm_ep_zcpy_recv_cancel();
 void test_efa_rdm_ep_post_handshake_error_handling_pke_exhaustion();