OSHMEM/SMPL/UCX: Add real fence support

+ Add quiet method to SPML, so it can have different implementation with
fence.
+ Use ucp_worker_fence for spml_fence method of UCX SPML

Signed-off-by: Mikhail Brinskii <mikhailb@mellanox.com>

oshmem/mca/scoll/basic/scoll_basic_alltoall.c

@@ -72,10 +72,10 @@ int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
        return rc;
     }
 
-    /* fence (which currently acts as quiet) is needed
+    /* quiet is needed because scoll level barrier does not
-     * because scoll level barrier does not guarantee put completion
+     * guarantee put completion
      */
-    MCA_SPML_CALL(fence());
+    MCA_SPML_CALL(quiet());
 
     /* Wait for operation completion */
     SCOLL_VERBOSE(14, "[#%d] Wait for operation completion", group->my_pe);

oshmem/mca/scoll/basic/scoll_basic_barrier.c

@@ -167,8 +167,7 @@ static int _algorithm_central_counter(struct oshmem_group_t *group,
          The root could leave the first barrier and in the second barrier it could get SHMEM_SYNC_WAIT value on
          remote node before the remote node receives its SHMEM_SYNC_RUN value in the first barrier
          */
-        /* TODO: actually it must be quiet */
+        MCA_SPML_CALL(quiet());
-        MCA_SPML_CALL(fence());
     }
     /* Wait for RUN signal */
     else {

oshmem/mca/scoll/basic/scoll_basic_broadcast.c

@@ -146,10 +146,10 @@ static int _algorithm_central_counter(struct oshmem_group_t *group,
                 rc = MCA_SPML_CALL(put(target, nlong, (void *)source, pe_cur));
             }
         }
-        /* fence (which currently acts as quiet) is needed
+        /* quiet is needed because scoll level barrier does not
-         * because scoll level barrier does not guarantee put completion 
+         * guarantee put completion
          */
-        MCA_SPML_CALL(fence());
+        MCA_SPML_CALL(quiet());
     }
 
     if (rc == OSHMEM_SUCCESS) {

oshmem/mca/spml/base/spml_base.c

@@ -153,11 +153,7 @@ int mca_spml_base_wait(void* addr, int cmp, void* value, int datatype)
  */
 int mca_spml_base_wait_nb(void* handle)
 {
-    /* TODO fence is a gag for more accurate code
+    MCA_SPML_CALL(quiet());
-     * Use shmem_quiet() (or a function calling shmem_quiet()) or
-     * shmem_wait_nb() to force completion of transfers for non-blocking operations.
-     */
-    MCA_SPML_CALL(fence());
 
     return OSHMEM_SUCCESS;
 }

oshmem/mca/spml/ikrit/spml_ikrit.c

@@ -168,6 +168,7 @@ mca_spml_ikrit_t mca_spml_ikrit = {
         mca_spml_ikrit_send,
         mca_spml_base_wait,
         mca_spml_base_wait_nb,
+        mca_spml_ikrit_fence, /* fence is implemented as quiet */
         mca_spml_ikrit_fence,
         mca_spml_ikrit_cache_mkeys,
         mca_spml_base_rmkey_free,

oshmem/mca/spml/spml.h

@@ -275,12 +275,19 @@ typedef int (*mca_spml_base_module_send_fn_t)(void *buf,
                                               mca_spml_base_put_mode_t mode);
 
 /**
- * Wait for completion of all outstanding put() requests
+ * Assures ordering of delivery of put() requests
  *
  * @return         - OSHMEM_SUCCESS or failure status.
  */
 typedef int (*mca_spml_base_module_fence_fn_t)(void);
 
+/**
+ * Wait for completion of all outstanding put() requests
+ *
+ * @return         - OSHMEM_SUCCESS or failure status.
+ */
+typedef int (*mca_spml_base_module_quiet_fn_t)(void);
+
 /**
  * Waits for completion of a non-blocking put or get issued by the calling PE.
  *
@@ -321,6 +328,7 @@ struct mca_spml_base_module_1_0_0_t {
     mca_spml_base_module_wait_fn_t spml_wait;
     mca_spml_base_module_wait_nb_fn_t spml_wait_nb;
     mca_spml_base_module_fence_fn_t spml_fence;
+    mca_spml_base_module_quiet_fn_t spml_quiet;
 
     mca_spml_base_module_mkey_unpack_fn_t spml_rmkey_unpack;
     mca_spml_base_module_mkey_free_fn_t   spml_rmkey_free;

oshmem/mca/spml/ucx/spml_ucx.c

@@ -60,8 +60,8 @@ mca_spml_ucx_t mca_spml_ucx = {
         mca_spml_ucx_send,
         mca_spml_base_wait,
         mca_spml_base_wait_nb,
-        mca_spml_ucx_quiet, /* At the moment fence is the same as quite for 
+        mca_spml_ucx_fence,
-                               every spml */
+        mca_spml_ucx_quiet,
         mca_spml_ucx_rmkey_unpack,
         mca_spml_ucx_rmkey_free,
         mca_spml_ucx_rmkey_ptr,
@@ -520,7 +520,7 @@ int mca_spml_ucx_deregister(sshmem_mkey_t *mkeys)
     spml_ucx_mkey_t   *ucx_mkey;
     map_segment_t *mem_seg;
 
-    MCA_SPML_CALL(fence());
+    MCA_SPML_CALL(quiet());
     if (!mkeys)
         return OSHMEM_SUCCESS;
 
@@ -598,7 +598,7 @@ int mca_spml_ucx_fence(void)
 {
     ucs_status_t err;
 
-    err = ucp_worker_flush(mca_spml_ucx.ucp_worker);
+    err = ucp_worker_fence(mca_spml_ucx.ucp_worker);
     if (UCS_OK != err) {
          SPML_ERROR("fence failed: %s", ucs_status_string(err));
          oshmem_shmem_abort(-1);

oshmem/shmem/c/shmem_barrier.c

@@ -36,7 +36,7 @@ void shmem_barrier(int PE_start, int logPE_stride, int PE_size, long *pSync)
 
 #if OSHMEM_SPEC_COMPAT == 1
     /* all outstanding puts must be completed */
-    shmem_fence();
+    shmem_quiet();
 #endif
 
     /* Create group basing PE_start, logPE_stride and PE_size */
@@ -54,7 +54,7 @@ void shmem_barrier_all(void)
 
 #if OSHMEM_SPEC_COMPAT == 1
     /* all outstanding puts must be completed */
-    shmem_fence();
+    shmem_quiet();
 #endif
 
     if (mca_scoll_sync_array) {

oshmem/shmem/c/shmem_quiet.c

@@ -23,5 +23,5 @@
 void shmem_quiet(void)
 {
 
-    MCA_SPML_CALL(fence());
+    MCA_SPML_CALL(quiet());
 }

oshmem/shmem/fortran/shmem_quiet_f.c

@@ -30,5 +30,5 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
 
 void shmem_quiet_f(void)
 {
-    MCA_SPML_CALL(fence());
+    MCA_SPML_CALL(quiet());
 }