prov/efa: Implement FI_OPT_SHARED_MEMORY_PERMITTED

This commit implements FI_OPT_SHARED_MEMORY_PERMITTED for the EFA
provider. All SHM provider resources are disabled when
FI_OPT_SHARED_MEMORY_PERMITTED is set to true.

Signed-off-by: Sai Sunku <[email protected]>

man/fi_efa.7.md

@@ -205,6 +205,12 @@ These OFI runtime parameters apply only to the RDM endpoint.
   [`ptrace protection`](https://wiki.ubuntu.com/SecurityTeam/Roadmap/KernelHardening#ptrace_Protection)
   is turned on. You can turn it off to enable shm transfer.
 
+  FI_EFA_ENABLE_SHM_TRANSFER is parsed during the fi_domain call and is related to the FI_OPT_SHARED_MEMORY_PERMITTED endpoint option.
+  If FI_EFA_ENABLE_SHM_TRANSFER is set to true, the FI_OPT_SHARED_MEMORY_PERMITTED endpoint
+  option overrides FI_EFA_ENABLE_SHM_TRANSFER. If FI_EFA_ENABLE_SHM_TRANSFER is set to false,
+  but the FI_OPT_SHARED_MEMORY_PERMITTED is set to true, the FI_OPT_SHARED_MEMORY_PERMITTED
+  setopt call will fail with -FI_EINVAL.
+
 *FI_EFA_SHM_AV_SIZE*
 : Defines the maximum number of entries in SHM provider's address vector.
 

prov/efa/src/rdm/efa_rdm_ep.h

@@ -215,6 +215,7 @@ struct efa_rdm_ep {
 	int	hmem_p2p_opt; /* what to do for hmem transfers */
 	struct fid_ep *peer_srx_ep; /* support sharing receive context with peer providers */
 	bool cuda_api_permitted; /**< whether end point is permitted to call CUDA API */
+	bool shm_permitted; /* Whether the endpoint is allowed to use shared memory for intra-node communication */
 
 	/* use_device_rdma:
 	   Can be set via fi_setopt in API >= 1.18.

prov/efa/src/rdm/efa_rdm_ep_fiops.c

@@ -461,6 +461,7 @@ int efa_rdm_ep_open(struct fid_domain *domain, struct fi_info *info,
 	efa_rdm_ep->efa_max_outstanding_rx_ops = efa_domain->device->rdm_info->rx_attr->size;
 	efa_rdm_ep->efa_device_iov_limit = efa_domain->device->rdm_info->tx_attr->iov_limit;
 	efa_rdm_ep->use_device_rdma = efa_rdm_get_use_device_rdma(info->fabric_attr->api_version);
+	efa_rdm_ep->shm_permitted = true;
 
 	cq_attr.size = MAX(efa_rdm_ep->rx_size + efa_rdm_ep->tx_size,
 			   efa_env.cq_size);
@@ -999,9 +1000,10 @@ void efa_rdm_ep_update_shm(struct efa_rdm_ep *ep)
 	 * AWS Neuron and Habana Synapse, have no SHM provider
 	 * support anyways, so disabling SHM will not impact them.
 	 */
-	if ((ep->user_info->caps & FI_HMEM)
+	if (((ep->user_info->caps & FI_HMEM)
 	    && hmem_ops[FI_HMEM_CUDA].initialized
-	    && !ep->cuda_api_permitted) {
+	    && !ep->cuda_api_permitted)
+		|| !ep->shm_permitted) {
 		use_shm = false;
 	}
 
@@ -1175,6 +1177,35 @@ static int efa_rdm_ep_set_cuda_api_permitted(struct efa_rdm_ep *ep, bool cuda_ap
 	return 0;
 }
 
+/**
+ * @brief act on shared_memory_permitted flag called by efa_rdm_ep_setopt
+ * @param[in,out]	ep			endpoint
+ * @param[in]		shm_permitted	whether shared memory is permitted
+ * @return		0 on success,
+ *			-FI_EINVAL if shm is requested but the FI_EFA_ENABLE_SHM_TRANSFER environment variable is set to false
+ * @related efa_rdm_ep
+ */
+static int efa_rdm_ep_set_shared_memory_permitted(struct efa_rdm_ep *ep, bool shm_permitted)
+{
+	if (!shm_permitted) {
+		EFA_WARN(FI_LOG_EP_CTRL,
+			 "FI_OPT_SHARED_MEMORY_PERMITTED set to false");
+		ep->shm_permitted = false;
+		return FI_SUCCESS;
+	}
+
+	if (!efa_env.enable_shm_transfer) {
+		EFA_WARN(FI_LOG_EP_CTRL,
+			 "FI_OPT_SHARED_MEMORY_PERMITTED endpoint option set "
+			 "to true but FI_EFA_ENABLE_SHM_TRANSFER environment "
+			 "variable is set to false.");
+		return -FI_EINVAL;
+	}
+
+	ep->shm_permitted = true;
+	return 0;
+}
+
 /**
  * @brief set use_device_rdma flag in efa_rdm_ep.
  *
@@ -1365,6 +1396,13 @@ static int efa_rdm_ep_setopt(fid_t fid, int level, int optname,
 		if (ret)
 			return ret;
 		break;
+	case FI_OPT_SHARED_MEMORY_PERMITTED:
+		if (optlen != sizeof(bool))
+			return -FI_EINVAL;
+		ret = efa_rdm_ep_set_shared_memory_permitted(efa_rdm_ep, *(bool *)optval);
+		if (ret)
+			return ret;
+		break;
 	case FI_OPT_EFA_USE_DEVICE_RDMA:
 		if (optlen != sizeof(bool))
 			return -FI_EINVAL;

prov/efa/test/efa_unit_test_ep.c

@@ -581,3 +581,24 @@ void test_efa_rdm_ep_getopt_oversized_optlen(struct efa_resource **state)
 {
 	test_efa_rdm_ep_getopt(state, 16, FI_SUCCESS);
 }
+
+void test_efa_rdm_ep_setopt_shared_memory_permitted(struct efa_resource **state)
+{
+	struct efa_resource *resource = *state;
+	struct efa_rdm_ep *ep;
+	bool optval = false;
+
+	efa_unit_test_resource_construct_ep_not_enabled(resource, FI_EP_RDM);
+
+	ep = container_of(resource->ep, struct efa_rdm_ep,
+			  base_ep.util_ep.ep_fid);
+
+	assert_int_equal(fi_setopt(&resource->ep->fid, FI_OPT_ENDPOINT,
+				   FI_OPT_SHARED_MEMORY_PERMITTED, &optval,
+				   sizeof(optval)),
+			 0);
+
+	assert_int_equal(fi_enable(resource->ep), 0);
+
+	assert_null(ep->shm_ep);
+}

prov/efa/test/efa_unit_tests.c

@@ -83,6 +83,7 @@ int main(void)
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_handshake_receive_without_peer_host_id_and_do_not_send_local_host_id, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_getopt_undersized_optlen, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_getopt_oversized_optlen, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
+		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_setopt_shared_memory_permitted, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_cq_create_error_handling, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_pkt_pool_flags, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),
 		cmocka_unit_test_setup_teardown(test_efa_rdm_ep_pkt_pool_page_alignment, efa_unit_test_mocks_setup, efa_unit_test_mocks_teardown),

prov/efa/test/efa_unit_tests.h

@@ -102,6 +102,7 @@ void test_efa_rdm_ep_dc_atomic_error_handling();
 void test_efa_rdm_ep_send_with_shm_no_copy();
 void test_efa_rdm_ep_rma_without_caps();
 void test_efa_rdm_ep_atomic_without_caps();
+void test_efa_rdm_ep_setopt_shared_memory_permitted();
 void test_dgram_cq_read_empty_cq();
 void test_ibv_cq_ex_read_empty_cq();
 void test_ibv_cq_ex_read_failed_poll();