Merge pull request #9975 from bwbarrett/bugfix/rdma-osc

osc/rdma: Handle remote completion behavior

Author: bwbarrett

ompi/mca/osc/rdma/Makefile.am

@@ -11,6 +11,8 @@
 # Copyright (c) 2014-2015 Los Alamos National Security, LLC. All rights
 #                         reserved.
 # Copyright (c) 2017      IBM Corporation.  All rights reserved.
+# Copyright (c) 2022      Amazon.com, Inc. or its affiliates.
+#                         All Rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -21,6 +23,8 @@
 rdma_sources = \
 	osc_rdma.h \
 	osc_rdma_module.c \
+	osc_rdma_btl_comm.h \
+	osc_rdma_btl_comm.c \
 	osc_rdma_comm.h \
 	osc_rdma_comm.c \
 	osc_rdma_accumulate.c \

ompi/mca/osc/rdma/osc_rdma.h

@@ -44,6 +44,7 @@
 #include "ompi/mca/osc/osc.h"
 #include "ompi/mca/osc/base/base.h"
 #include "opal/mca/btl/btl.h"
+#include "opal/mca/btl/base/btl_base_am_rdma.h"
 #include "ompi/memchecker.h"
 #include "ompi/op/op.h"
 #include "opal/align.h"
@@ -57,8 +58,6 @@
 
 #define RANK_ARRAY_COUNT(module) ((ompi_comm_size ((module)->comm) + (module)->node_count - 1) / (module)->node_count)
 
-#define MCA_OSC_RDMA_BTLS_SIZE_INIT 4
-
 enum {
     OMPI_OSC_RDMA_LOCKING_TWO_LEVEL,
     OMPI_OSC_RDMA_LOCKING_ON_DEMAND,
@@ -150,9 +149,6 @@ struct ompi_osc_rdma_module_t {
     /** value of same_size info key for this window */
     bool same_size;
 
-    /** CPU atomics can be used */
-    bool use_cpu_atomics;
-
     /** passive-target synchronization will not be used in this window */
     bool no_locks;
 
@@ -260,18 +256,38 @@ struct ompi_osc_rdma_module_t {
     /** lock for peer hash table/array */
     opal_mutex_t peer_lock;
 
+    /* ******************* communication *********************** */
+
+    /* we currently support two modes of operation, a single
+     * accelerated btl (which can use memory registration and can use
+     * btl_flush() and one or more alternate btls, which cannot use
+     * flush() or rely on memory registration.  Since it is an
+     * either/or situation, we use a union to simplify the code.
+     */
+    bool use_accelerated_btl;
+
+    union {
+        struct {
+            mca_btl_base_module_t *accelerated_btl;
+        };
+        struct {
+            mca_btl_base_am_rdma_module_t **alternate_am_rdmas;
+            uint8_t alternate_btl_count;
+        };
+    };
+
+    /** Does the selected BTL require memory registration?  This field
+        will be false when alternate BTLs are used, and the value
+        when an accelerated BTL is used depends on the registration
+        requirements of the underlying BTL. */
+    bool use_memory_registration;
 
-    /** BTL(s) in use. Currently this is only used to support RDMA emulation over
-     * non-RDMA BTLs. The typical usage is btl/sm + btl/tcp. In the future this
-     * could be used to support multiple RDMA-capable BTLs but the memory registration
-     * paths will need to be updated to pack/unpack multiple registration handles. */
-    struct mca_btl_base_module_t **selected_btls;
-    uint8_t selected_btls_size;
-    uint8_t btls_in_use;
+    size_t put_alignment;
+    size_t get_alignment;
+    size_t put_limit;
+    size_t get_limit;
 
-    /** Only true if one BTL is in use. Memory registration is only supported when
-     * using a single BTL. */
-    bool use_memory_registration;
+    uint32_t atomic_flags;
 
     /** registered fragment used for locally buffered RDMA transfers */
     struct ompi_osc_rdma_frag_t *rdma_frag;
@@ -383,10 +399,11 @@ static inline int _ompi_osc_rdma_register (ompi_osc_rdma_module_t *module, struc
                                            size_t size, uint32_t flags, mca_btl_base_registration_handle_t **handle, int line, const char *file)
 {
     if (module->use_memory_registration) {
+        assert(module->use_accelerated_btl);
         OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_INFO, "registering segment with btl. range: %p - %p (%lu bytes)",
                          ptr, (void*)((char *) ptr + size), size);
 
-        *handle = module->selected_btls[0]->btl_register_mem (module->selected_btls[0], endpoint, ptr, size, flags);
+        *handle = module->accelerated_btl->btl_register_mem(module->accelerated_btl, endpoint, ptr, size, flags);
         if (OPAL_UNLIKELY(NULL == *handle)) {
             OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_DEBUG, "failed to register pointer with selected BTL. base: %p, "
                              "size: %lu. file: %s, line: %d", ptr, (unsigned long) size, file, line);
@@ -404,7 +421,9 @@ static inline int _ompi_osc_rdma_register (ompi_osc_rdma_module_t *module, struc
 static inline void _ompi_osc_rdma_deregister (ompi_osc_rdma_module_t *module, mca_btl_base_registration_handle_t *handle, int line, const char *file)
 {
     if (handle) {
-        module->selected_btls[0]->btl_deregister_mem (module->selected_btls[0], handle);
+        assert(module->use_memory_registration);
+        assert(module->use_accelerated_btl);
+        module->accelerated_btl->btl_deregister_mem(module->accelerated_btl, handle);
     }
 }
 
@@ -536,10 +555,11 @@ static inline ompi_osc_rdma_sync_t *ompi_osc_rdma_module_sync_lookup (ompi_osc_r
 static bool ompi_osc_rdma_use_btl_flush (ompi_osc_rdma_module_t *module)
 {
 #if defined(BTL_VERSION) && (BTL_VERSION >= 310)
-    return !!(module->selected_btls[0]->btl_flush);
-#else
-    return false;
+    if (module->use_accelerated_btl) {
+        return (NULL != module->accelerated_btl->btl_flush);
+    }
 #endif
+    return false;
 }
 
 /**
@@ -601,13 +621,13 @@ static inline void ompi_osc_rdma_sync_rdma_complete (ompi_osc_rdma_sync_t *sync)
         opal_progress ();
     }  while (ompi_osc_rdma_sync_get_count (sync));
 #else
-    mca_btl_base_module_t *btl_module = sync->module->selected_btls[0];
-
     do {
         if (!ompi_osc_rdma_use_btl_flush (sync->module)) {
             opal_progress ();
         } else {
-            btl_module->btl_flush (btl_module, NULL);
+            assert(sync->module->use_accelerated_btl);
+            mca_btl_base_module_t *btl_module = sync->module->accelerated_btl;
+            btl_module->btl_flush(btl_module, NULL);
         }
     }  while (ompi_osc_rdma_sync_get_count (sync) || (sync->module->rdma_frag && (sync->module->rdma_frag->pending > 1)));
 #endif
@@ -637,17 +657,20 @@ static inline bool ompi_osc_rdma_oor (int rc)
 
 __opal_attribute_always_inline__
 static inline mca_btl_base_module_t *ompi_osc_rdma_selected_btl (ompi_osc_rdma_module_t *module, uint8_t btl_index) {
-    return module->selected_btls[btl_index];
+    if (module->use_accelerated_btl) {
+        assert(0 == btl_index);
+        return module->accelerated_btl;
+    } else {
+        assert(btl_index < module->alternate_btl_count);
+        return module->alternate_am_rdmas[btl_index]->btl;
+    }
 }
 
-__opal_attribute_always_inline__
-static inline void ompi_osc_rdma_selected_btl_insert (ompi_osc_rdma_module_t *module, struct mca_btl_base_module_t *btl, uint8_t btl_index) {
-    if(btl_index == module->selected_btls_size) {
-        module->selected_btls_size *= 2;
-        module->selected_btls = realloc(module->selected_btls, module->selected_btls_size * sizeof(struct mca_btl_base_module_t *));
-        assert(NULL != module->selected_btls);
-    }
-    module->selected_btls[btl_index] = btl;
+
+static inline mca_btl_base_am_rdma_module_t *ompi_osc_rdma_selected_am_rdma(ompi_osc_rdma_module_t *module, uint8_t btl_index) {
+    assert(!module->use_accelerated_btl);
+    assert(btl_index < module->alternate_btl_count);
+    return module->alternate_am_rdmas[btl_index];
 }
 
 #endif /* OMPI_OSC_RDMA_H */

ompi/mca/osc/rdma/osc_rdma_accumulate.c

@@ -10,17 +10,24 @@
  * Copyright (c) 2019-2021 Google, LLC. All rights reserved.
  * Copyright (c) 2021      IBM Corporation.  All rights reserved.
  * Copyright (c) 2022      Cisco Systems, Inc.  All rights reserved
+ * Copyright (c) 2022      Amazon.com, Inc. or its affiliates.
+ *                         All Rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
  *
  * $HEADER$
  */
 
+#include "ompi_config.h"
+
 #include "osc_rdma_accumulate.h"
 #include "osc_rdma_request.h"
 #include "osc_rdma_comm.h"
+#include "osc_rdma_lock.h"
+#include "osc_rdma_btl_comm.h"
 
+#include "opal/util/minmax.h"
 #include "ompi/mca/osc/base/base.h"
 #include "ompi/mca/osc/base/osc_base_obj_convert.h"
 
@@ -157,13 +164,11 @@ static int ompi_osc_rdma_fetch_and_op_atomic (ompi_osc_rdma_sync_t *sync, const
                                               mca_btl_base_registration_handle_t *target_handle, ompi_op_t *op, ompi_osc_rdma_request_t *req)
 {
     ompi_osc_rdma_module_t *module = sync->module;
-    mca_btl_base_module_t *selected_btl = ompi_osc_rdma_selected_btl (module, peer->data_btl_index);
-    int32_t atomic_flags = selected_btl->btl_atomic_flags;
     int btl_op, flags;
     int64_t origin;
 
-    if ((8 != extent && !((MCA_BTL_ATOMIC_SUPPORTS_32BIT & atomic_flags) && 4 == extent)) ||
-        (!(OMPI_DATATYPE_FLAG_DATA_INT & dt->super.flags) && !(MCA_BTL_ATOMIC_SUPPORTS_FLOAT & atomic_flags)) ||
+    if ((8 != extent && !((MCA_BTL_ATOMIC_SUPPORTS_32BIT & module->atomic_flags) && 4 == extent)) ||
+        (!(OMPI_DATATYPE_FLAG_DATA_INT & dt->super.flags) && !(MCA_BTL_ATOMIC_SUPPORTS_FLOAT & module->atomic_flags)) ||
         !ompi_op_is_intrinsic (op) || (0 == ompi_osc_rdma_op_mapping[op->op_type])) {
         return OMPI_ERR_NOT_SUPPORTED;
     }
@@ -235,19 +240,11 @@ static int ompi_osc_rdma_acc_single_atomic (ompi_osc_rdma_sync_t *sync, const vo
                                             ompi_op_t *op, ompi_osc_rdma_request_t *req)
 {
     ompi_osc_rdma_module_t *module = sync->module;
-    mca_btl_base_module_t *selected_btl = ompi_osc_rdma_selected_btl (module, peer->data_btl_index);
-    int32_t atomic_flags = selected_btl->btl_atomic_flags;
     int btl_op, flags;
     int64_t origin;
 
-    if (!(selected_btl->btl_flags & MCA_BTL_FLAGS_ATOMIC_OPS)) {
-        /* btl put atomics not supported or disabled. fall back on fetch-and-op */
-        return ompi_osc_rdma_fetch_and_op_atomic (sync, origin_addr, NULL, dt, extent, peer, target_address, target_handle,
-                                                  op, req);
-    }
-
-    if ((8 != extent && !((MCA_BTL_ATOMIC_SUPPORTS_32BIT & atomic_flags) && 4 == extent)) ||
-        (!(OMPI_DATATYPE_FLAG_DATA_INT & dt->super.flags) && !(MCA_BTL_ATOMIC_SUPPORTS_FLOAT & atomic_flags)) ||
+    if ((8 != extent && !((MCA_BTL_ATOMIC_SUPPORTS_32BIT & module->atomic_flags) && 4 == extent)) ||
+        (!(OMPI_DATATYPE_FLAG_DATA_INT & dt->super.flags) && !(MCA_BTL_ATOMIC_SUPPORTS_FLOAT & module->atomic_flags)) ||
         !ompi_op_is_intrinsic (op) || (0 == ompi_osc_rdma_op_mapping[op->op_type])) {
         return OMPI_ERR_NOT_SUPPORTED;
     }
@@ -585,9 +582,9 @@ static inline int ompi_osc_rdma_gacc_master (ompi_osc_rdma_sync_t *sync, const v
 
             /* determine how much to put in this operation */
             if (source_count) {
-                acc_len = min(min(target_iovec[target_iov_index].iov_len, source_iovec[source_iov_index].iov_len), acc_limit);
+                acc_len = opal_min(opal_min(target_iovec[target_iov_index].iov_len, source_iovec[source_iov_index].iov_len), acc_limit);
             } else {
-                acc_len = min(target_iovec[target_iov_index].iov_len, acc_limit);
+                acc_len = opal_min(target_iovec[target_iov_index].iov_len, acc_limit);
             }
 
             if (0 != acc_len) {
@@ -662,13 +659,11 @@ static inline int ompi_osc_rdma_cas_atomic (ompi_osc_rdma_sync_t *sync, const vo
                                             bool lock_acquired)
 {
     ompi_osc_rdma_module_t *module = sync->module;
-    mca_btl_base_module_t *btl = ompi_osc_rdma_selected_btl (module, peer->data_btl_index);
-    int32_t atomic_flags = btl->btl_atomic_flags;
     const size_t size = datatype->super.size;
     int64_t compare, source;
     int flags, ret;
 
-    if (8 != size && !(4 == size && (MCA_BTL_ATOMIC_SUPPORTS_32BIT & atomic_flags))) {
+    if (8 != size && !(4 == size && (MCA_BTL_ATOMIC_SUPPORTS_32BIT & module->atomic_flags))) {
         return OMPI_ERR_NOT_SUPPORTED;
     }
 
@@ -716,7 +711,6 @@ static inline int cas_rdma (ompi_osc_rdma_sync_t *sync, const void *source_addr,
                             mca_btl_base_registration_handle_t *target_handle, bool lock_acquired)
 {
     ompi_osc_rdma_module_t *module = sync->module;
-    mca_btl_base_module_t *btl = ompi_osc_rdma_selected_btl (module, peer->data_btl_index);
     unsigned long len = datatype->super.size;
     mca_btl_base_registration_handle_t *local_handle = NULL;
     ompi_osc_rdma_frag_t *frag = NULL;
@@ -741,26 +735,30 @@ static inline int cas_rdma (ompi_osc_rdma_sync_t *sync, const void *source_addr,
         return OMPI_SUCCESS;
     }
 
-    if (btl->btl_register_mem && len > btl->btl_put_local_registration_threshold) {
-        do {
-            ret = ompi_osc_rdma_frag_alloc (module, len, &frag, &ptr);
-            if (OPAL_UNLIKELY(OMPI_SUCCESS == ret)) {
-                break;
-            }
+    if (module->use_memory_registration) {
+        mca_btl_base_module_t *btl = ompi_osc_rdma_selected_btl (module, peer->data_btl_index);
+        if (len > btl->btl_put_local_registration_threshold) {
+            do {
+                ret = ompi_osc_rdma_frag_alloc(module, len, &frag, &ptr);
+                if (OPAL_UNLIKELY(OMPI_SUCCESS == ret)) {
+                    break;
+                }
 
-            ompi_osc_rdma_progress (module);
-        } while (1);
+                ompi_osc_rdma_progress (module);
+            } while (1);
 
-        memcpy (ptr, source_addr, len);
-        local_handle = frag->handle;
+            memcpy(ptr, source_addr, len);
+            local_handle = frag->handle;
+        }
     }
 
     OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "RDMA compare-and-swap initiating blocking btl put...");
 
     do {
-        ret = btl->btl_put (btl, peer->data_endpoint, ptr, target_address,
-                            local_handle, target_handle, len, 0, MCA_BTL_NO_ORDER,
-                            ompi_osc_rdma_cas_put_complete, (void *) &complete, NULL);
+        ret = ompi_osc_rdma_btl_put(module, peer->data_btl_index, peer->data_endpoint,
+                                    ptr, target_address, local_handle, target_handle,
+                                    len, 0, MCA_BTL_NO_ORDER,
+                                    ompi_osc_rdma_cas_put_complete, (void *) &complete, NULL);
         if (OPAL_SUCCESS == ret || (OPAL_ERR_OUT_OF_RESOURCE != ret && OPAL_ERR_TEMP_OUT_OF_RESOURCE != ret)) {
             break;
         }

ompi/mca/osc/rdma/osc_rdma_active_target.c

@@ -77,33 +77,6 @@ OBJ_CLASS_INSTANCE(ompi_osc_rdma_pending_op_t, opal_list_item_t,
                    ompi_osc_rdma_pending_op_construct,
                    ompi_osc_rdma_pending_op_destruct);
 
-/**
- * Dummy completion function for atomic operations
- */
-void ompi_osc_rdma_atomic_complete (mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
-                                    void *local_address, mca_btl_base_registration_handle_t *local_handle,
-                                    void *context, void *data, int status)
-{
-    ompi_osc_rdma_pending_op_t *pending_op = (ompi_osc_rdma_pending_op_t *) context;
-
-    OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_INFO, "pending atomic %p complete with status %d", (void*)pending_op, status);
-
-    if (pending_op->op_result) {
-        memmove (pending_op->op_result, pending_op->op_buffer, pending_op->op_size);
-    }
-
-    if (NULL != pending_op->cbfunc) {
-        pending_op->cbfunc (pending_op->cbdata, pending_op->cbcontext, status);
-    }
-
-    if (NULL != pending_op->op_frag) {
-        ompi_osc_rdma_frag_complete (pending_op->op_frag);
-        pending_op->op_frag = NULL;
-    }
-
-    pending_op->op_complete = true;
-    OBJ_RELEASE(pending_op);
-}
 
 /**
  * compare_ranks: