REVIEW: various fixes for DPU

openucx · Feb 19, 2025 · 41f0b4e · 41f0b4e
1 parent 9452315
commit 41f0b4e
Show file tree

Hide file tree

Showing 16 changed files with 367 additions and 132 deletions.
diff --git a/src/components/cl/basic/cl_basic_context.c b/src/components/cl/basic/cl_basic_context.c
@@ -61,20 +61,20 @@ UCC_CLASS_CLEANUP_FUNC(ucc_cl_basic_context_t)
     ucc_free(self->super.tl_ctxs);
 }
 
-ucc_status_t ucc_cl_basic_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_cl_basic_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                   void *address, size_t len, void *memh,
                                   void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_cl_basic_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_cl_basic_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                     void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_cl_basic_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_cl_basic_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                     void *memh, void **packed_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/cl/hier/cl_hier_context.c b/src/components/cl/hier/cl_hier_context.c
@@ -92,20 +92,20 @@ ucc_cl_hier_get_context_attr(const ucc_base_context_t *context, /* NOLINT */
     return UCC_OK;
 }
 
-ucc_status_t ucc_cl_hier_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_cl_hier_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                  void *address, size_t len, void *memh,
                                  void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_cl_hier_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_cl_hier_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                    void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_cl_hier_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_cl_hier_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **packed_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/tl/cuda/tl_cuda_context.c b/src/components/tl/cuda/tl_cuda_context.c
@@ -76,19 +76,19 @@ UCC_CLASS_INIT_FUNC(ucc_tl_cuda_context_t,
     return status;
 }
 
-ucc_status_t ucc_tl_cuda_mem_map(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_cuda_mem_map(const ucc_base_context_t *context, /* NOLINT */
                                  void *address, size_t len, void *memh)
 {
     return UCC_ERR_NOT_IMPLEMENTED;
 }
 
-ucc_status_t ucc_tl_cuda_mem_unmap(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_cuda_mem_unmap(const ucc_base_context_t *context, /* NOLINT */
                                    void                     *memh)
 {
     return UCC_ERR_NOT_IMPLEMENTED;
 }
 
-ucc_status_t ucc_tl_cuda_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_cuda_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_IMPLEMENTED;

diff --git a/src/components/tl/mlx5/tl_mlx5_context.c b/src/components/tl/mlx5/tl_mlx5_context.c
@@ -303,20 +303,20 @@ ucc_status_t ucc_tl_mlx5_context_create_epilog(ucc_base_context_t *context)
     return ucc_tl_mlx5_context_ib_ctx_pd_setup(context);
 }
 
-ucc_status_t ucc_tl_mlx5_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_mlx5_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                  void *address, size_t len, void *memh,
                                  void *tl_h)
 {
     return UCC_ERR_NOT_IMPLEMENTED;
 }
 
-ucc_status_t ucc_tl_mlx5_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_mlx5_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                    void *memh)
 {
     return UCC_ERR_NOT_IMPLEMENTED;
 }
 
-ucc_status_t ucc_tl_mlx5_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_mlx5_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_IMPLEMENTED;

diff --git a/src/components/tl/nccl/tl_nccl_context.c b/src/components/tl/nccl/tl_nccl_context.c
@@ -224,20 +224,20 @@ ucc_tl_nccl_get_context_attr(const ucc_base_context_t *context, /* NOLINT */
     return UCC_OK;
 }
 
-ucc_status_t ucc_tl_nccl_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_nccl_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                  void *address, size_t len, void *memh,
                                  void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_nccl_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_nccl_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                    void *memh)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_nccl_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_nccl_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/tl/rccl/tl_rccl_context.c b/src/components/tl/rccl/tl_rccl_context.c
@@ -121,20 +121,20 @@ ucc_tl_rccl_get_context_attr(const ucc_base_context_t *context, /* NOLINT */
     return UCC_OK;
 }
 
-ucc_status_t ucc_tl_rccl_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_rccl_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                  void *address, size_t len, void *memh,
                                  void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_rccl_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_rccl_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                    void *memh)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_rccl_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_rccl_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/tl/self/tl_self_context.c b/src/components/tl/self/tl_self_context.c
@@ -50,20 +50,20 @@ ucc_tl_self_get_context_attr(const ucc_base_context_t *context, /* NOLINT */
     return UCC_OK;
 }
 
-ucc_status_t ucc_tl_self_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_self_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                  void *address, size_t len, void *memh,
                                  void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_self_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_self_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                    void *memh)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_self_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_self_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                    void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/tl/sharp/tl_sharp_context.c b/src/components/tl/sharp/tl_sharp_context.c
@@ -512,20 +512,20 @@ ucc_status_t ucc_tl_sharp_get_context_attr(const ucc_base_context_t *context, /*
     return UCC_OK;
 }
 
-ucc_status_t ucc_tl_sharp_mem_map(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_sharp_mem_map(const ucc_base_context_t *context, int type, /* NOLINT */
                                   void *address, size_t len, void *memh,
                                   void *tl_h)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_sharp_mem_unmap(const ucc_base_context_t *context, int type,
+ucc_status_t ucc_tl_sharp_mem_unmap(const ucc_base_context_t *context, int type, /* NOLINT */
                                     void *memh)
 {
     return UCC_ERR_NOT_SUPPORTED;
 }
 
-ucc_status_t ucc_tl_sharp_memh_pack(const ucc_base_context_t *context,
+ucc_status_t ucc_tl_sharp_memh_pack(const ucc_base_context_t *context, /* NOLINT */
                                     void *memh, void **pack_buffer)
 {
     return UCC_ERR_NOT_SUPPORTED;

diff --git a/src/components/tl/ucp/alltoall/alltoall_onesided.c b/src/components/tl/ucp/alltoall/alltoall_onesided.c
@@ -35,18 +35,18 @@ ucc_status_t ucc_tl_ucp_alltoall_onesided_start(ucc_coll_task_t *ctask)
 
     UCPCHECK_GOTO(
         ucc_tl_ucp_put_nb((void *)(src + start * nelems), (void *)dest, nelems,
-                          start, src_memh[start], dst_memh[start], team, task),
+                          start, *src_memh, dst_memh[start], team, task),
                           task, out);
-    UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, start, src_memh[start],
+    UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, start, *src_memh,
                                         dst_memh[start], team),
                                         task, out);
 
     for (peer = (start + 1) % gsize; peer != start; peer = (peer + 1) % gsize) {
         UCPCHECK_GOTO(ucc_tl_ucp_put_nb(
                           (void *)(src + peer * nelems), (void *)dest, nelems,
-                          peer, src_memh[peer], dst_memh[peer], team, task),
+                          peer, *src_memh, dst_memh[peer], team, task),
                           task, out);
-        UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, peer, src_memh[peer],
+        UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, peer, *src_memh,
                                             dst_memh[peer], team),
                                             task, out);
     }

diff --git a/src/components/tl/ucp/alltoallv/alltoallv_onesided.c b/src/components/tl/ucp/alltoallv/alltoallv_onesided.c
@@ -48,11 +48,10 @@ ucc_status_t ucc_tl_ucp_alltoallv_onesided_start(ucc_coll_task_t *ctask)
 
         UCPCHECK_GOTO(ucc_tl_ucp_put_nb(PTR_OFFSET(src, sd_disp),
                                         PTR_OFFSET(dest, dd_disp),
-                                        data_size, peer, src_memh[peer],
+                                        data_size, peer, *src_memh,
                                         dst_memh[peer], team, task),
                       task, out);
-        UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, peer, src_memh[peer],
-                                            dst_memh[peer], team), task, out);
+        UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, peer, *src_memh, dst_memh[peer], team), task, out);
     }
     return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);
 out:

diff --git a/src/components/tl/ucp/tl_ucp.h b/src/components/tl/ucp/tl_ucp.h
@@ -197,6 +197,24 @@ extern ucc_config_field_t ucc_tl_ucp_lib_config_table[];
 #define UCC_TL_UCP_REMOTE_RKEY(_ctx, _rank, _seg)                              \
     ((_ctx)->rkeys[_rank * _ctx->n_rinfo_segs + _seg])
 
+#define UCC_TL_UCP_MEMH_TL_HEADERS        2
+
+#define UCC_TL_UCP_MEMH_TL_PACKED_HEADERS 2
+
+#define UCC_TL_UCP_MEMH_TL_HEADER_SIZE                                         \
+    (sizeof(size_t) * UCC_TL_UCP_MEMH_TL_HEADERS)
+
+#define UCC_TL_UCP_MEMH_TL_PACKED_HEADER_SIZE                                  \
+    (sizeof(size_t) *                                                          \
+     (UCC_TL_UCP_MEMH_TL_HEADERS + UCC_TL_UCP_MEMH_TL_PACKED_HEADERS))
+
+#define UCC_TL_UCP_MEMH_TL_KEY_SIZE(buffer)                                    \
+    (*(size_t *)(PTR_OFFSET(buffer, UCC_TL_UCP_MEMH_TL_HEADER_SIZE)))
+
+#define UCC_TL_UCP_MEMH_TL_PACKED_MEMH(buffer)                                 \
+    (PTR_OFFSET(buffer, UCC_TL_UCP_MEMH_TL_PACKED_HEADER_SIZE +                \
+                            UCC_TL_UCP_MEMH_TL_KEY_SIZE(buffer)))
+
 extern ucs_memory_type_t ucc_memtype_to_ucs[UCC_MEMORY_TYPE_LAST+1];
 
 void ucc_tl_ucp_pre_register_mem(ucc_tl_ucp_team_t *team, void *addr,