linux: don't hide the NVIDIA GPU node on non-POWER platforms

And allow to force hide it with HWLOC_KEEP_NVIDIA_GPU_NUMA_NODES=0.

These nodes were hidden by default on POWER because people
could use interleaved memory allocations across sockets,
but they would also interleave across GPU memory.

On NVIDIA Grace Hopper, interleaved allocation isn't much
of an issue since there's a single CPU node (and 1 single GPU
node per GPU slice).

Thanks to Antoine Morvan for the report.

Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>
(cherry picked from commit 39fae7e)

NEWS

@@ -28,6 +28,7 @@ Version 2.10.1
   thanks to Florent Pruvost for the help.
 * Fix the enabling of CUDA in Windows CMake build,
   Thanks to Moritz Kreutzer for the patch.
+* Don't hide the GPU NUMA node on NVIDIA Grace Hopper.
 
 
 Version 2.10.0

doc/hwloc.doxy

@@ -1320,10 +1320,13 @@ following environment variables.
 
 <dt>HWLOC_KEEP_NVIDIA_GPU_NUMA_NODES=0</dt>
   <dd>show or hide NUMA nodes that correspond to NVIDIA GPU memory.
-  By default they are ignored to avoid interleaved memory being allocated
-  on GPU by mistake.
-  Setting this environment variable to 1 exposes these NUMA nodes.
-  They may be recognized by the <em>GPUMemory</em> subtype.
+  By default they are ignored on POWER platforms to avoid interleaved
+  memory being allocated on GPU by mistake.
+
+  Setting this environment variable to 0 hides the NUMA nodes (default on POWER).
+  Setting to 1 exposes these NUMA nodes (default on non-POWER platforms such as NVIDIA Grace Hopper).
+
+  These NUMA nodes may be recognized by the <em>GPUMemory</em> subtype.
   They also have a <em>PCIBusID</em> info attribute to identify the
   corresponding GPU.
   </dd>

hwloc/topology-linux.c

@@ -1,6 +1,6 @@
 /*
  * Copyright © 2009 CNRS
- * Copyright © 2009-2023 Inria.  All rights reserved.
+ * Copyright © 2009-2024 Inria.  All rights reserved.
  * Copyright © 2009-2013, 2015, 2020 Université Bordeaux
  * Copyright © 2009-2018 Cisco Systems, Inc.  All rights reserved.
  * Copyright © 2015 Intel, Inc.  All rights reserved.
@@ -4242,7 +4242,10 @@ look_sysfsnode(struct hwloc_topology *topology,
 	struct dirent *dirent;
 	int keep;
 	env = getenv("HWLOC_KEEP_NVIDIA_GPU_NUMA_NODES");
-        keep = env && atoi(env);
+        /* NVIDIA GPU NUMA nodes hidden by default on POWER */
+        keep = (data->arch != HWLOC_LINUX_ARCH_POWER);
+        if (env)
+          keep = atoi(env);
 	while ((dirent = readdir(dir)) != NULL) {
 	  char nvgpunumapath[300], line[256];
           int err;