MCA/COMMON/UCX: minor unification of del_proces calls

- some common functionality of del_procs calls is moved into
  mca_common module
- blocking ucp_put call is replaced by non-blocking routine

Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>

ompi/mca/pml/ucx/pml_ucx.c

@@ -18,6 +18,7 @@
 #include "opal/mca/pmix/pmix.h"
 #include "ompi/message/message.h"
 #include "ompi/mca/pml/base/pml_base_bsend.h"
+#include "opal/mca/common/ucx/common_ucx.h"
 #include "pml_ucx_request.h"
 
 #include <inttypes.h>
@@ -374,29 +375,19 @@ static void mca_pml_ucx_waitall(void **reqs, int *count_p)
 
     PML_UCX_VERBOSE(2, "waiting for %d disconnect requests", *count_p);
     for (i = 0; i < *count_p; ++i) {
-        do {
-            opal_progress();
-            status = ucp_request_test(reqs[i], NULL);
-        } while (status == UCS_INPROGRESS);
+        status = opal_common_ucx_wait_request(reqs[i], ompi_pml_ucx.ucp_worker);
         if (status != UCS_OK) {
             PML_UCX_ERROR("disconnect request failed: %s",
                           ucs_status_string(status));
         }
-        ucp_request_free(reqs[i]);
         reqs[i] = NULL;
     }
 
     *count_p = 0;
 }
 
-static void mca_pml_fence_complete_cb(int status, void *fenced)
-{
-    *(int*)fenced = 1;
-}
-
 int mca_pml_ucx_del_procs(struct ompi_proc_t **procs, size_t nprocs)
 {
-    volatile int fenced = 0;
     ompi_proc_t *proc;
     int num_reqs;
     size_t max_reqs;
@@ -447,10 +438,7 @@ int mca_pml_ucx_del_procs(struct ompi_proc_t **procs, size_t nprocs)
     mca_pml_ucx_waitall(dreqs, &num_reqs);
     free(dreqs);
 
-    opal_pmix.fence_nb(NULL, 0, mca_pml_fence_complete_cb, (void*)&fenced);
-    while (!fenced) {
-        ucp_worker_progress(ompi_pml_ucx.ucp_worker);
-    }
+    opal_common_ucx_mca_pmix_fence(ompi_pml_ucx.ucp_worker);
 
     return OMPI_SUCCESS;
 }

opal/mca/common/ucx/common_ucx.c

@@ -11,6 +11,7 @@
 
 #include "common_ucx.h"
 #include "opal/mca/base/mca_base_var.h"
+#include "opal/mca/pmix/pmix.h"
 
 /***********************************************************************/
 
@@ -36,3 +37,19 @@ OPAL_DECLSPEC void opal_common_ucx_mca_register(void)
 void opal_common_ucx_empty_complete_cb(void *request, ucs_status_t status)
 {
 }
+
+static void opal_common_ucx_mca_fence_complete_cb(int status, void *fenced)
+{
+    *(int*)fenced = 1;
+}
+
+OPAL_DECLSPEC void opal_common_ucx_mca_pmix_fence(ucp_worker_h worker)
+{
+    volatile int fenced = 0;
+
+    opal_pmix.fence_nb(NULL, 0, opal_common_ucx_mca_fence_complete_cb, (void*)&fenced);
+    while (!fenced) {
+        ucp_worker_progress(worker);
+    }
+}
+

opal/mca/common/ucx/common_ucx.h

@@ -27,6 +27,7 @@ extern int opal_common_ucx_progress_iterations;
 
 OPAL_DECLSPEC void opal_common_ucx_mca_register(void);
 OPAL_DECLSPEC void opal_common_ucx_empty_complete_cb(void *request, ucs_status_t status);
+OPAL_DECLSPEC void opal_common_ucx_mca_pmix_fence(ucp_worker_h worker);
 
 static inline
 ucs_status_t opal_common_ucx_wait_request(ucs_status_ptr_t request, ucp_worker_h worker)

oshmem/mca/spml/ucx/spml_ucx.c

@@ -103,15 +103,11 @@ static void mca_spml_ucx_waitall(void **reqs, int *count_p)
 
     SPML_VERBOSE(10, "waiting for %d disconnect requests", *count_p);
     for (i = 0; i < *count_p; ++i) {
-        do {
-            opal_progress();
-            status = ucp_request_test(reqs[i], NULL);
-        } while (status == UCS_INPROGRESS);
+        status = opal_common_ucx_wait_request(reqs[i], mca_spml_ucx.ucp_worker);
         if (status != UCS_OK) {
             SPML_ERROR("disconnect request failed: %s",
                        ucs_status_string(status));
         }
-        ucp_request_release(reqs[i]);
         reqs[i] = NULL;
     }
 
@@ -175,8 +171,9 @@ int mca_spml_ucx_del_procs(ompi_proc_t** procs, size_t nprocs)
     mca_spml_ucx_waitall(dreqs, &num_reqs);
     free(dreqs);
 
-    opal_pmix.fence(NULL, 0);
+    opal_common_ucx_mca_pmix_fence(mca_spml_ucx.ucp_worker);
     free(mca_spml_ucx.ucp_peers);
+    mca_spml_ucx.ucp_peers = NULL;
     return OSHMEM_SUCCESS;
 }
 
@@ -585,12 +582,14 @@ int mca_spml_ucx_put(void* dst_addr, size_t size, void* src_addr, int dst)
 {
     void *rva;
     ucs_status_t status;
+    ucs_status_ptr_t request;
     spml_ucx_mkey_t *ucx_mkey;
 
     ucx_mkey = mca_spml_ucx_get_mkey(dst, dst_addr, &rva, &mca_spml_ucx);
-    status = ucp_put(mca_spml_ucx.ucp_peers[dst].ucp_conn, src_addr, size,
-                     (uint64_t)rva, ucx_mkey->rkey);
-
+    request = ucp_put_nb(mca_spml_ucx.ucp_peers[dst].ucp_conn, src_addr, size,
+                         (uint64_t)rva, ucx_mkey->rkey, opal_common_ucx_empty_complete_cb);
+    /* TODO: replace wait_request by opal_common_ucx_wait_request_opal_status */
+    status = opal_common_ucx_wait_request(request, mca_spml_ucx.ucp_worker);
     return ucx_status_to_oshmem(status);
 }