[PATCH i-g-t v2 6/6] tests/amdgpu: query KFD aperture node count before filling array

[Junhua Shen <Junhua.Shen@amd.com>]

kfd_get_gpuvm_base() passed num_of_nodes = NUM_OF_SUPPORTED_GPUS
unconditionally, which the kernel rejects with -EINVAL via
kfd_ioctl_get_process_apertures_new()'s
'num_of_nodes > kfd_topology_get_num_devices()' check.

Follow the documented two-step protocol instead: an initial call with
num_of_nodes == 0 returns the actual node count without touching the
array, then fill up to that count (capped at the array capacity).

Signed-off-by: Junhua Shen <Junhua.Shen@amd.com>
---
 tests/amdgpu/amd_kfd_dmabuf_unload.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/tests/amdgpu/amd_kfd_dmabuf_unload.c b/tests/amdgpu/amd_kfd_dmabuf_unload.c
index 28caf8982..f0d076910 100644
--- a/tests/amdgpu/amd_kfd_dmabuf_unload.c
+++ b/tests/amdgpu/amd_kfd_dmabuf_unload.c
@@ -132,9 +132,15 @@ static int kfd_get_gpuvm_base(int fd, uint32_t gpu_id,
 	struct kfd_ioctl_get_process_apertures_new_args args = {};
 	uint32_t i;
 
+	if (ioctl(fd, AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, &args))
+		return -errno;
+	if (args.num_of_nodes == 0)
+		return -ENODEV;
+	if (args.num_of_nodes > NUM_OF_SUPPORTED_GPUS)
+		args.num_of_nodes = NUM_OF_SUPPORTED_GPUS;
+
 	args.kfd_process_device_apertures_ptr =
 		(uint64_t)(uintptr_t)aps;
-	args.num_of_nodes = NUM_OF_SUPPORTED_GPUS;
 	if (ioctl(fd, AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, &args))
 		return -errno;
 
-- 
2.34.1