prov/cxi: Define FI_CXI_FORCE_DEV_REG_COPY

FI_CXI_FORCE_DEV_REG_COPY will force the CXI provider to use the HMEM device register copy routines. If not supported, RDMA operations or memory registration will fail. Signed-off-by: Ian Ziemba <[email protected]>
ofiwg · Nov 20, 2024 · 7943a1b · 7943a1b
1 parent 3372122
commit 7943a1b
Show file tree

Hide file tree

Showing 4 changed files with 17 additions and 3 deletions.
diff --git a/man/fi_cxi.7.md b/man/fi_cxi.7.md
@@ -1301,6 +1301,10 @@ The CXI provider checks for the following environment variables:
 :   Max amount of time to poll when LE invalidate disabling an MR configured with MR
     match events.
 
+*FI_CXI_FORCE_DEV_REG_COPY*
+:   Force the CXI provider to use the HMEM device register copy routines. If not
+    supported, RDMA operations or memory registration will fail.
+
 Note: Use the fi_info utility to query provider environment variables:
 <code>fi_info -p cxi -e</code>
 

diff --git a/prov/cxi/include/cxip.h b/prov/cxi/include/cxip.h
@@ -324,6 +324,7 @@ struct cxip_environment {
 	int hybrid_unexpected_msg_preemptive;
 	size_t mr_cache_events_disable_poll_nsecs;
 	size_t mr_cache_events_disable_le_poll_nsecs;
+	int force_dev_reg_copy;
 };
 
 extern struct cxip_environment cxip_env;

diff --git a/prov/cxi/src/cxip_ep.c b/prov/cxi/src/cxip_ep.c
@@ -1187,8 +1187,10 @@ int cxip_ep_setopt_priv(struct cxip_ep *ep, int level, int optname,
 		if (!cuda_api_permitted && !cuda_is_gdrcopy_enabled())
 			return -FI_EOPNOTSUPP;
 
-		ep->ep_obj->require_dev_reg_copy[FI_HMEM_CUDA] =
-			!cuda_api_permitted;
+		if (!cxip_env.force_dev_reg_copy) {
+			ep->ep_obj->require_dev_reg_copy[FI_HMEM_CUDA] =
+				!cuda_api_permitted;
+		}
 		break;
 
 	default:
@@ -1296,7 +1298,7 @@ int cxip_alloc_endpoint(struct cxip_domain *cxip_dom, struct fi_info *hints,
 	 * disables it.
 	 */
 	for (i = 0; i < OFI_HMEM_MAX; i++)
-		ep_obj->require_dev_reg_copy[i] = false;
+		ep_obj->require_dev_reg_copy[i] = cxip_env.force_dev_reg_copy;
 
 	ofi_atomic_initialize32(&ep_obj->txq_ref, 0);
 	ofi_atomic_initialize32(&ep_obj->tgq_ref, 0);

diff --git a/prov/cxi/src/cxip_info.c b/prov/cxi/src/cxip_info.c
@@ -670,6 +670,7 @@ struct cxip_environment cxip_env = {
 		CXIP_MR_CACHE_EVENTS_DISABLE_POLL_NSECS,
 	.mr_cache_events_disable_le_poll_nsecs =
 		CXIP_MR_CACHE_EVENTS_DISABLE_LE_POLL_NSECS,
+	.force_dev_reg_copy = false,
 };
 
 static void cxip_env_init(void)
@@ -1288,6 +1289,12 @@ static void cxip_env_init(void)
 	fi_param_get_size_t(&cxip_prov, "mr_cache_events_disable_le_poll_nsecs",
 			    &cxip_env.mr_cache_events_disable_le_poll_nsecs);
 
+	fi_param_define(&cxip_prov, "force_dev_reg_copy", FI_PARAM_BOOL,
+			"Force device register copy operations. Default: %d",
+			cxip_env.force_dev_reg_copy);
+	fi_param_get_bool(&cxip_prov, "force_dev_reg_copy",
+			  &cxip_env.force_dev_reg_copy);
+
 	set_system_page_size();
 }