feat(rdma): constrain C linkage to init

On builds where lttng is enabled, the inclusion of nccl_ofi_tracepoint.h
pulls in <lttng/ust-utils.h>, which defines templates when __cplusplus
is defined. Those templates cannot have C linkage.

This file does not really need nccl_ofi_tracepoint.h; it only needs
config.h to test for whether NVTX tracing is enabled and nvtx3 headers
for conditionally defining the nvtx domain handles.

It's also simultaneously true that the large majority of these headers
are consumed only by rdma.c, so another option is to just move the guard
down to where it's actually required.

stack-info: PR: aws#591, branch: aws-nslick/stack/32
Signed-off-by: Nicholas Sielicki <[email protected]>

include/nccl_ofi_rdma.h

@@ -5,9 +5,7 @@
 #ifndef NCCL_OFI_RDMA_H_
 #define NCCL_OFI_RDMA_H_
 
-#ifdef __cplusplus
-extern "C" {
-#endif
+#include "config.h"
 
 #include <uthash/uthash.h>
 #include <rdma/fabric.h>
@@ -21,7 +19,9 @@ extern "C" {
 #include "nccl_ofi_msgbuff.h"
 #include "nccl_ofi_scheduler.h"
 #include "nccl_ofi_topo.h"
-#include "nccl_ofi_tracepoint.h"
+#if HAVE_NVTX_TRACING
+#include <nvtx3/nvToolsExt.h>
+#endif
 
 /* Maximum number of rails supported. This defines the size of
  * messages exchanged during connection establishment (linear
@@ -840,6 +840,9 @@ typedef struct nccl_net_ofi_rdma_device {
 #endif
 } nccl_net_ofi_rdma_device_t;
 
+#ifdef __cplusplus
+extern "C" {
+#endif
 
 struct nccl_net_ofi_rdma_plugin {
 	nccl_net_ofi_plugin_t base;