osc/rdma: Split OMPI_OSC_RDMA_PEER_CPU_ATOMICS from OMPI_OSC_RDMA_PEER_LOCAL_BASE

We need to distinguish between the ability of using cpu atomics and
being able to access the peer's memory.

Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>

Author: devreal

ompi/mca/osc/rdma/osc_rdma_accumulate.c

@@ -990,7 +990,7 @@ int ompi_osc_rdma_compare_and_swap (const void *origin_addr, const void *compare
      * OR if we have an exclusive lock
      * OR if other processes won't try to use the network either */
     bool use_shared_mem = module->single_node ||
-                          (ompi_osc_rdma_peer_local_base (peer) &&
+                          (ompi_osc_rdma_peer_cpu_atomics (peer) &&
                               (ompi_osc_rdma_peer_is_exclusive (peer) ||
                                   !module->acc_single_intrinsic));
 
@@ -1013,7 +1013,7 @@ int ompi_osc_rdma_compare_and_swap (const void *origin_addr, const void *compare
         lock_acquired = true;
     }
 
-    if (ompi_osc_rdma_peer_local_base (peer)) {
+    if (ompi_osc_rdma_peer_cpu_atomics (peer)) {
         ret = ompi_osc_rdma_cas_local (origin_addr, compare_addr, result_addr, dt,
                                        peer, target_address, target_handle, module,
                                        lock_acquired);
@@ -1095,7 +1095,7 @@ int ompi_osc_rdma_rget_accumulate_internal (ompi_win_t *win, const void *origin_
         (void) ompi_osc_rdma_lock_acquire_exclusive (module, peer, offsetof (ompi_osc_rdma_state_t, accumulate_lock));
     }
 
-    if (ompi_osc_rdma_peer_local_base (peer)) {
+    if (ompi_osc_rdma_peer_cpu_atomics (peer)) {
         /* local/self optimization */
         ret = ompi_osc_rdma_gacc_local (origin_addr, origin_count, origin_datatype, result_addr, result_count,
                                         result_datatype, peer, target_address, target_handle, target_count,

ompi/mca/osc/rdma/osc_rdma_component.c

@@ -530,6 +530,7 @@ static int allocate_state_single (ompi_osc_rdma_module_t *module, void **base, s
     module->my_peer = my_peer;
     module->free_after = module->rank_array;
     my_peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
+    my_peer->flags |= OMPI_OSC_RDMA_PEER_CPU_ATOMICS;
     my_peer->state = (uint64_t) (uintptr_t) module->state;
 
     if (use_cpu_atomics) {
@@ -839,6 +840,7 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
             if (MPI_WIN_FLAVOR_DYNAMIC == module->flavor) {
                 if (use_cpu_atomics && peer_rank == my_rank) {
                     peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
+                    peer->flags |= OMPI_OSC_RDMA_PEER_CPU_ATOMICS;
                 }
                 /* nothing more to do */
                 continue;
@@ -853,7 +855,7 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
                 ex_peer->size = temp[i].size;
             }
 
-            if (use_cpu_atomics && (MPI_WIN_FLAVOR_ALLOCATE == module->flavor || peer_rank == my_rank)) {
+            if (MPI_WIN_FLAVOR_ALLOCATE == module->flavor || peer_rank == my_rank) {
                 /* base is local and cpu atomics are available */
                 if (MPI_WIN_FLAVOR_ALLOCATE == module->flavor) {
                     ex_peer->super.base = (uintptr_t) module->segment_base + offset;
@@ -862,6 +864,9 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
                 }
 
                 peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
+                if (use_cpu_atomics) {
+                    peer->flags |= OMPI_OSC_RDMA_PEER_CPU_ATOMICS;
+                }
                 offset += temp[i].size;
             } else {
                 ex_peer->super.base = peer_region->base;

ompi/mca/osc/rdma/osc_rdma_peer.h

@@ -142,6 +142,8 @@ enum {
     OMPI_OSC_RDMA_PEER_BASE_FREE            = 0x40,
     /** peer was demand locked as part of lock-all (when in demand locking mode) */
     OMPI_OSC_RDMA_PEER_DEMAND_LOCKED        = 0x80,
+    /** we can use CPU atomics on that peer */
+    OMPI_OSC_RDMA_PEER_CPU_ATOMICS          = 0x100,
 };
 
 /**
@@ -224,6 +226,11 @@ static inline bool ompi_osc_rdma_peer_local_base (ompi_osc_rdma_peer_t *peer)
     return !!(peer->flags & OMPI_OSC_RDMA_PEER_LOCAL_BASE);
 }
 
+static inline bool ompi_osc_rdma_peer_cpu_atomics (ompi_osc_rdma_peer_t *peer)
+{
+    return ompi_osc_rdma_peer_local_base(peer) && !!(peer->flags & OMPI_OSC_RDMA_PEER_CPU_ATOMICS);
+}
+
 /**
  * @brief check if the peer's state pointer is local to this process
  *