UCT/GDAKI: Fix wqe_idx overflow (#10881)

itayalroy · web-flow · commit 9ed46592ea2c · 2025-09-15T10:37:23.000+03:00
uct_gdaki_wqe_prepare_put_or_atomic assumes
wqe_idx is at most UINT16_T_MAX but atomic &amp; single
APIs passed sq_rsvd_index, an ever-incrementing
uint64_t counter.

Other submit APIs avoid this issue by using inc_mask() per lane,
which properly masks the WQE index down to uint16_t.

Signed-off-by: Itay Alroy &lt;ialroy@nvidia.com&gt;
diff --git a/AUTHORS b/AUTHORS
@@ -46,6 +46,7 @@ Hui Zhou <hzhou321@anl.gov>
 Igor Ivanov <igori@mellanox.com>
 Ilia Yastrebov <iyastrebov@nvidia.com>
 Ilya Nelkenbaum <ilyan@mellanox.com>
+Itay Alroy <ialroy@nvidia.com>
 Ivan Kochin <ikochin@nvidia.com>
 Jakir Kham <jakirkham@gmail.com>
 Jason Gunthorpe <jgg@mellanox.com>
diff --git a/src/uct/ib/mlx5/gdaki/gdaki.cuh b/src/uct/ib/mlx5/gdaki/gdaki.cuh
@@ -109,7 +109,7 @@ uct_rc_mlx5_gda_reserv_wqe(struct doca_gpu_dev_verbs_qp *qp, unsigned count,
 
 UCS_F_DEVICE void uct_rc_mlx5_gda_wqe_prepare_put_or_atomic(
         doca_gpu_dev_verbs_qp *qp, doca_gpu_dev_verbs_wqe *wqe_ptr,
-        uint32_t wqe_idx, uint32_t opcode, unsigned ctrl_flags, uint64_t raddr,
+        uint16_t wqe_idx, uint32_t opcode, unsigned ctrl_flags, uint64_t raddr,
         uint32_t rkey, uint64_t laddr, uint32_t lkey, uint32_t bytes,
         bool is_atomic, uint64_t add)
 {

Original file line number	Diff line number	Diff line change
`@@ -109,7 +109,7 @@ uct_rc_mlx5_gda_reserv_wqe(struct doca_gpu_dev_verbs_qp *qp, unsigned count,`
`109`	`109`
`110`	`110`	`UCS_F_DEVICE void uct_rc_mlx5_gda_wqe_prepare_put_or_atomic(`
`111`	`111`	`doca_gpu_dev_verbs_qp qp, doca_gpu_dev_verbs_wqe wqe_ptr,`
`112`		`- uint32_t wqe_idx, uint32_t opcode, unsigned ctrl_flags, uint64_t raddr,`
	`112`	`+ uint16_t wqe_idx, uint32_t opcode, unsigned ctrl_flags, uint64_t raddr,`
`113`	`113`	`uint32_t rkey, uint64_t laddr, uint32_t lkey, uint32_t bytes,`
`114`	`114`	`bool is_atomic, uint64_t add)`
`115`	`115`	`{`