[ISSUE]: Uninitialized ah_attr in ibgda_create_dct_shared_objects leads to potential EINVAL

[jcao219]

nvshmem/src/modules/transport/ibgda/ibgda.cpp

Line 2336 in 2d7d25f

struct ibv_ah_attr ah_attr;

Within the ibgda_create_dct_shared_objects function, the struct ibv_ah_attr ah_attr variable appears to be used without being initialized.

Specifically, the problematic field is ah_attr.static_rate then containing a garbage value. This uninitialized data is then used later when calling ibv_create_ah, which I've seen leading to an EINVAL (Invalid argument) error on a RoCE setup.

Steps to Reproduce

1. Build nvshmem with IBGDA transport enabled.
2. Run a simple application with nvshmem_init, such as the hello.cpp example with NVSHMEM_IB_ENABLE_IBGDA=1

[benhg]

Thank you for the report. We will address it in the next release of NVSHMEM.

[cclinuxer]

`
diff --git a/src/modules/transport/ibgda/ibgda.cpp b/src/modules/transport/ibgda/ibgda.cpp
index 4c69cce..4aaaea6 100644
--- a/src/modules/transport/ibgda/ibgda.cpp
+++ b/src/modules/transport/ibgda/ibgda.cpp
@@ -2382,6 +2382,7 @@ static int ibgda_create_dct_shared_objects(nvshmemt_ibgda_state_t *ibgda_state,
NVSHMEMI_NULL_ERROR_JMP(recv_cq, status, NVSHMEMX_ERROR_INTERNAL, out,
"ibv_create_cq for recv_cq failed.\n");

• memset(&ah_attr, 0, sizeof(ah_attr));
  if (port_attr->lid == 0) {
  ib_get_gid_index(&ftable, device->common_device.context, portid, port_attr->gid_tbl_len,
  (int *)&device->common_device.gid_info[portid - 1].local_gid_index,
  `

[cclinuxer]

May I ask when the next NVSHMEM release is scheduled?

[benhg]

NVSHMEM 3.5.19 was released on Wednesday, so I will close this issue.