Merge pull request #13330 from devreal/win-shared-query

janjust · web-flow · commit 333d8ad94fca · 2025-10-06T14:21:29.000-05:00
MPI 4.0: Allow MPI_WIN_SHARED_QUERY on regular windows
diff --git a/docs/man-openmpi/man3/MPI_Win_shared_query.3.rst b/docs/man-openmpi/man3/MPI_Win_shared_query.3.rst
@@ -28,19 +28,24 @@ DESCRIPTION
 
 :ref:`MPI_Win_shared_query` queries the process-local address for
 remote memory segments created with
-:ref:`MPI_Win_allocate_shared`. This function can return different
+:ref:`MPI_Win_create`, :ref:`MPI_Win_allocate`, or
+:ref:`MPI_Win_allocate_shared`. Shared memory is only guaranteed to be
+available for windows created with :ref:`MPI_Win_allocate_shared`.
+This function can return different
 process-local addresses for the same physical memory on different
 processes. The returned memory can be used for load/store accesses
-subject to the constraints defined in MPI-3.1 section 11.7. This
-function can only be called with windows of flavor
-MPI_WIN_FLAVOR_SHARED. If the passed window is not of flavor
-MPI_WIN_FLAVOR_SHARED, the error MPI_ERR_RMA_FLAVOR is raised. When
+subject to the constraints defined in MPI-3.1 section 11.7.
+If the passed a dynamic window, the error MPI_ERR_RMA_FLAVOR is raised. When
 rank is ``MPI_PROC_NULL``, the *pointer*, *disp_unit*, and *size* returned
 are the pointer, disp_unit, and size of the memory segment belonging
-the lowest rank that specified *size* > 0. If all processes in the
+to the lowest rank that specified *size* > 0. If all processes in the
 group attached to the window specified *size* = 0, then the call
 returns *size* = 0 and a *baseptr* as if :ref:`MPI_Alloc_mem` was
 called with *size* = 0.
+If the window was created with :ref:`MPI_Win_create` or
+:ref:`MPI_Win_allocate` and shared memory is not supported for these windows,
+the call returns *size* = 0 and a *baseptr* as if
+:ref:`MPI_Alloc_mem` was called with *size* = 0.
 
 
 C NOTES
@@ -59,3 +64,5 @@ ERRORS
 .. seealso::
    * :ref:`MPI_Alloc_mem`
    * :ref:`MPI_Win_allocate_shared`
+   * :ref:`MPI_Win_create`
+   * :ref:`MPI_Win_allocate`
diff --git a/ompi/mca/osc/rdma/osc_rdma_accumulate.c b/ompi/mca/osc/rdma/osc_rdma_accumulate.c
@@ -13,6 +13,7 @@
  * Copyright (c) 2022      Amazon.com, Inc. or its affiliates.
  *                         All Rights reserved.
  * Copyright (c) 2023      Jeffrey M. Squyres.  All rights reserved.
+ * Copyright (c) 2025      Stony Brook University.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -990,7 +991,7 @@ 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 +1014,7 @@ 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 +1096,7 @@ 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,
diff --git a/ompi/mca/osc/rdma/osc_rdma_component.c b/ompi/mca/osc/rdma/osc_rdma_component.c
@@ -25,6 +25,7 @@
  * Copyright (c) 2020-2021 Google, LLC. All rights reserved.
  * Copyright (c) 2019-2021 Triad National Security, LLC. All rights
  *                         reserved.
+ * Copyright (c) 2025      Stony Brook University.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -69,6 +70,8 @@
 #include "ompi/mca/bml/base/base.h"
 #include "ompi/mca/mtl/base/base.h"
 
+static int ompi_osc_rdma_shared_query(struct ompi_win_t *win, int rank, size_t *size,
+                               ptrdiff_t *disp_unit, void *baseptr);
 static int ompi_osc_rdma_component_register (void);
 static int ompi_osc_rdma_component_init (bool enable_progress_threads, bool enable_mpi_threads);
 static int ompi_osc_rdma_component_finalize (void);
@@ -113,6 +116,7 @@ ompi_osc_rdma_component_t mca_osc_rdma_component = {
 MCA_BASE_COMPONENT_INIT(ompi, osc, rdma)
 
 ompi_osc_base_module_t ompi_osc_rdma_module_rdma_template = {
+    .osc_win_shared_query = ompi_osc_rdma_shared_query,
     .osc_win_attach = ompi_osc_rdma_attach,
     .osc_win_detach  = ompi_osc_rdma_detach,
     .osc_free = ompi_osc_rdma_free,
@@ -527,6 +531,7 @@ 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) {
@@ -636,7 +641,6 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
     /* ensure proper alignment */
     if (MPI_WIN_FLAVOR_ALLOCATE == module->flavor) {
         data_base += OPAL_ALIGN_PAD_AMOUNT(data_base, memory_alignment);
-        size += OPAL_ALIGN_PAD_AMOUNT(size, memory_alignment);
     }
 
     do {
@@ -836,6 +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;
@@ -850,7 +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;
@@ -859,7 +864,11 @@ 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;
+                offset += OPAL_ALIGN_PAD_AMOUNT(offset, memory_alignment);
             } else {
                 ex_peer->super.base = peer_region->base;
 
@@ -898,7 +907,7 @@ static void ompi_osc_rdma_ensure_local_add_procs (void)
             /* this will cause add_proc to get called if it has not already been called */
             (void) mca_bml_base_get_endpoint (proc);
         }
-    } 
+    }
 
     free(procs);
 }
@@ -1632,3 +1641,60 @@ ompi_osc_rdma_set_no_lock_info(opal_infosubscriber_t *obj, const char *key, cons
      */
     return module->no_locks ? "true" : "false";
 }
+
+int ompi_osc_rdma_shared_query(
+    struct ompi_win_t *win, int rank, size_t *size,
+    ptrdiff_t *disp_unit, void *baseptr)
+{
+    int rc = OMPI_ERR_NOT_SUPPORTED;
+    ompi_osc_rdma_peer_t *peer;
+    int actual_rank = rank;
+    ompi_osc_rdma_module_t *module = GET_MODULE(win);
+
+    peer = ompi_osc_module_get_peer (module, actual_rank);
+    if (NULL == peer) {
+        return OMPI_ERR_NOT_SUPPORTED;
+    }
+
+    /* currently only supported for allocated windows */
+    if (MPI_WIN_FLAVOR_ALLOCATE != module->flavor) {
+        return OMPI_ERR_NOT_SUPPORTED;
+    }
+
+    if (!ompi_osc_rdma_peer_local_base(peer)) {
+        return OMPI_ERR_NOT_SUPPORTED;
+    }
+
+    if (MPI_PROC_NULL == rank) {
+        /* iterate until we find a rank that has a non-zero size */
+        for (int i = 0 ; i < ompi_comm_size(module->comm) ; ++i) {
+            peer = ompi_osc_module_get_peer (module, i);
+            ompi_osc_rdma_peer_extended_t *ex_peer = (ompi_osc_rdma_peer_extended_t *) peer;
+            if (!ompi_osc_rdma_peer_local_base(peer)) {
+                continue;
+            } else if (module->same_size && ex_peer->super.base) {
+                break;
+            } else if (ex_peer->size > 0) {
+                break;
+            }
+        }
+    }
+
+    if (module->same_size && module->same_disp_unit) {
+        *size = module->size;
+        *disp_unit = module->disp_unit;
+        ompi_osc_rdma_peer_basic_t *ex_peer = (ompi_osc_rdma_peer_basic_t *) peer;
+        *((void**) baseptr) = (void *) (intptr_t)ex_peer->base;
+        rc = OMPI_SUCCESS;
+    } else {
+        ompi_osc_rdma_peer_extended_t *ex_peer = (ompi_osc_rdma_peer_extended_t *) peer;
+        if (ex_peer->super.base != 0) {
+            /* we know the base of the peer */
+            *((void**) baseptr) = (void *) (intptr_t)ex_peer->super.base;
+            *size = ex_peer->size;
+            *disp_unit = ex_peer->disp_unit;
+            rc = OMPI_SUCCESS;
+        }
+    }
+    return rc;
+}
diff --git a/ompi/mca/osc/rdma/osc_rdma_peer.h b/ompi/mca/osc/rdma/osc_rdma_peer.h
@@ -2,6 +2,7 @@
 /*
  * Copyright (c) 2014-2018 Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2025      Stony Brook University.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -142,6 +143,8 @@ 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 +227,11 @@ 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
  *
diff --git a/ompi/mca/osc/sm/osc_sm_component.c b/ompi/mca/osc/sm/osc_sm_component.c
@@ -15,6 +15,7 @@
  * Copyright (c) 2020      FUJITSU LIMITED.  All rights reserved.
  * Copyright (c) 2020      High Performance Computing Center Stuttgart,
  *                         University of Stuttgart.  All rights reserved.
+ * Copyright (c) 2025      Stony Brook University.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -283,12 +284,12 @@ component_select(struct ompi_win_t *win, void **base, size_t size, ptrdiff_t dis
         if (module->noncontig) {
             opal_output_verbose(MCA_BASE_VERBOSE_DEBUG, ompi_osc_base_framework.framework_output,
                                 "allocating window using non-contiguous strategy");
-            total = ((size - 1) / pagesize + 1) * pagesize;
         } else {
             opal_output_verbose(MCA_BASE_VERBOSE_DEBUG, ompi_osc_base_framework.framework_output,
                                 "allocating window using contiguous strategy");
-            total = size;
         }
+
+        total = size;
         ret = module->comm->c_coll->coll_allgather(&total, 1, MPI_UNSIGNED_LONG,
                                                   rbuf, 1, MPI_UNSIGNED_LONG,
                                                   module->comm,
@@ -301,6 +302,9 @@ component_select(struct ompi_win_t *win, void **base, size_t size, ptrdiff_t dis
         total = 0;
         for (i = 0 ; i < comm_size ; ++i) {
             total += rbuf[i];
+            if (module->noncontig) {
+                total += OPAL_ALIGN_PAD_AMOUNT(total, pagesize);
+            }
         }
 
         /* user opal/shmem directly to create a shared memory segment */
@@ -378,6 +382,9 @@ component_select(struct ompi_win_t *win, void **base, size_t size, ptrdiff_t dis
             if (module->sizes[i] || !module->noncontig) {
                 module->bases[i] = ((char *) module->segment_base) + total;
                 total += rbuf[i];
+                if (module->noncontig) {
+                    total += OPAL_ALIGN_PAD_AMOUNT(total, pagesize);
+                }
             } else {
                 module->bases[i] = NULL;
             }
@@ -481,10 +488,6 @@ ompi_osc_sm_shared_query(struct ompi_win_t *win, int rank, size_t *size, ptrdiff
     ompi_osc_sm_module_t *module =
         (ompi_osc_sm_module_t*) win->w_osc_module;
 
-    if (module->flavor != MPI_WIN_FLAVOR_SHARED) {
-        return MPI_ERR_WIN;
-    }
-
     if (MPI_PROC_NULL != rank) {
         *size = module->sizes[rank];
         *((void**) baseptr) = module->bases[rank];
diff --git a/ompi/mca/osc/ucx/osc_ucx.h b/ompi/mca/osc/ucx/osc_ucx.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (C) Mellanox Technologies Ltd. 2001-2017. ALL RIGHTS RESERVED.
+ * Copyright (c) 2025      Stony Brook University.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -120,7 +121,8 @@ typedef struct ompi_osc_ucx_module {
     ompi_osc_base_module_t super;
     struct ompi_communicator_t *comm;
     int flavor;
-    size_t size;
+    size_t    size;
+    size_t   *sizes; /* used if not every process has the same size */
     uint64_t *addrs;
     uint64_t *state_addrs;
     uint64_t *comm_world_ranks;
@@ -143,13 +145,13 @@ typedef struct ompi_osc_ucx_module {
     bool lock_all_is_nocheck;
     bool no_locks;
     bool acc_single_intrinsic;
+    bool same_size;
     opal_common_ucx_ctx_t *ctx;
     opal_common_ucx_wpmem_t *mem;
     opal_common_ucx_wpmem_t *state_mem;
     ompi_osc_ucx_mem_ranges_t *epoc_outstanding_ops_mems;
     bool skip_sync_check;
     bool noncontig_shared_win;
-    size_t *sizes;
     /* in shared windows, shmem_addrs can be used for direct load store to
      * remote windows */
     uint64_t *shmem_addrs;
@@ -171,7 +173,7 @@ typedef struct ompi_osc_ucx_lock {
 } ompi_osc_ucx_lock_t;
 
 #define OSC_UCX_GET_EP(_module, rank_) (mca_osc_ucx_component.endpoints[_module->comm_world_ranks[rank_]])
-#define OSC_UCX_GET_DISP(module_, rank_) ((module_->disp_unit < 0) ? module_->disp_units[rank_] : module_->disp_unit)
+#define OSC_UCX_GET_DISP(module_, rank_) ompi_osc_ucx_get_disp_unit((module_), (rank_))
 
 #define OSC_UCX_GET_DEFAULT_EP(_ep_ptr, _module, _target)                   \
     if (opal_common_ucx_thread_enabled) {                  \
@@ -275,4 +277,24 @@ int ompi_osc_find_attached_region_position(ompi_osc_dynamic_win_info_t *dynamic_
 int ompi_osc_ucx_dynamic_lock(ompi_osc_ucx_module_t *module, int target);
 int ompi_osc_ucx_dynamic_unlock(ompi_osc_ucx_module_t *module, int target);
 
+/* returns the size at the peer */
+static inline size_t ompi_osc_ucx_get_size(ompi_osc_ucx_module_t *module, int rank)
+{
+    if (module->sizes) {
+        return module->sizes[rank];
+    } else {
+        return module->size;
+    }
+}
+
+/* returns the displacement unit for the given peer */
+static inline ptrdiff_t ompi_osc_ucx_get_disp_unit(ompi_osc_ucx_module_t *module, int rank)
+{
+    if (module->disp_units) {
+        return module->disp_units[rank];
+    } else {
+        return module->disp_unit;
+    }
+}
+
 #endif /* OMPI_OSC_UCX_H */
diff --git a/ompi/mca/osc/ucx/osc_ucx_component.c b/ompi/mca/osc/ucx/osc_ucx_component.c
diff --git a/ompi/mpi/c/win_shared_query.c.in b/ompi/mpi/c/win_shared_query.c.in
