Added support for openib RDMA READ.. note that performance is currently an

issue so PUT is default.. We are determining if this is an openib issue or a
btl issue as we have seen performance increases on mvapi. 

This commit was SVN r6928.

ompi/mca/btl/openib/btl_openib.c

@@ -32,7 +32,6 @@
 #include "mca/mpool/openib/mpool_openib.h" 
 #include <errno.h> 
 #include <string.h> 
-extern int errno; 
 
 mca_btl_openib_module_t mca_btl_openib_module = {
     {
@@ -57,7 +56,7 @@ mca_btl_openib_module_t mca_btl_openib_module = {
         mca_btl_openib_prepare_dst,
         mca_btl_openib_send,
         mca_btl_openib_put,
-        NULL /* get */ 
+        mca_btl_openib_get /* get */ 
     }
 };
 
@@ -775,7 +774,7 @@ int mca_btl_openib_send(
 }
 
 /*
- * RDMA local buffer to remote buffer address.
+ * RDMA WRITE local buffer to remote buffer address.
  */
 
 int mca_btl_openib_put( mca_btl_base_module_t* btl,
@@ -812,6 +811,45 @@ int mca_btl_openib_put( mca_btl_base_module_t* btl,
 
 }
 
+
+/*
+ * RDMA READ remote buffer to local buffer address.
+ */
+
+int mca_btl_openib_get( mca_btl_base_module_t* btl,
+                    mca_btl_base_endpoint_t* endpoint,
+                    mca_btl_base_descriptor_t* descriptor)
+{
+    struct ibv_send_wr* bad_wr; 
+    mca_btl_openib_frag_t* frag = (mca_btl_openib_frag_t*) descriptor; 
+    frag->endpoint = endpoint;
+    frag->wr_desc.sr_desc.opcode = IBV_WR_RDMA_READ; 
+    frag->wr_desc.sr_desc.send_flags = IBV_SEND_SIGNALED; 
+    frag->wr_desc.sr_desc.wr.rdma.remote_addr = (uintptr_t) frag->base.des_src->seg_addr.pval; 
+    frag->wr_desc.sr_desc.wr.rdma.rkey = frag->base.des_src->seg_key.key32[0]; 
+    frag->sg_entry.addr = (uintptr_t) frag->base.des_dst->seg_addr.pval; 
+    frag->sg_entry.length  = frag->base.des_dst->seg_len; 
+    
+    BTL_VERBOSE(("frag->wr_desc.sr_desc.wr.rdma.remote_addr = %llu .rkey = %lu frag->sg_entry.addr = %llu .length = %lu" 
+                  , frag->wr_desc.sr_desc.wr.rdma.remote_addr 
+                  , frag->wr_desc.sr_desc.wr.rdma.rkey
+                  , frag->sg_entry.addr
+                  , frag->sg_entry.length)); 
+
+    if(ibv_post_send(endpoint->lcl_qp_low, 
+                     &frag->wr_desc.sr_desc, 
+                     &bad_wr)){ 
+        BTL_ERROR(("error posting send request errno says %s", strerror(errno))); 
+        return OMPI_ERROR; 
+    }  
+    
+    MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(endpoint, 1); 
+    MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(endpoint, 1); 
+    
+    return OMPI_SUCCESS; 
+
+}
+
 /* 
  * Initialize the btl module by allocating a protection domain 
  *  and creating both the high and low priority completion queues 

ompi/mca/btl/openib/btl_openib.h

@@ -306,7 +306,24 @@ extern int mca_btl_openib_put(
     struct mca_btl_base_module_t* btl,
     struct mca_btl_base_endpoint_t* btl_peer,
     struct mca_btl_base_descriptor_t* decriptor
-);
+    );
+
+/**
+ * PML->BTL Initiate a get of the specified size.
+ *
+ * @param btl (IN)               BTL instance
+ * @param btl_base_peer (IN)     BTL peer addressing
+ * @param send_request (IN/OUT)  Send request (allocated by PML via mca_btl_base_request_alloc_fn_t)
+ * @param size (IN)              Number of bytes PML is requesting BTL to deliver
+ * @param flags (IN)             Flags that should be passed to the peer via the message header.
+ * @param request (OUT)          OMPI_SUCCESS if the BTL was able to queue one or more fragments
+ */
+extern int mca_btl_openib_get(
+    struct mca_btl_base_module_t* btl,
+    struct mca_btl_base_endpoint_t* btl_peer,
+    struct mca_btl_base_descriptor_t* decriptor
+    );
+    
 
 /**
  * Allocate a descriptor.

ompi/mca/btl/openib/btl_openib_component.c

@@ -41,7 +41,6 @@
 #include <errno.h> 
 #include <string.h>   /* for strerror()*/ 
 
-extern int errno; 
 mca_btl_openib_component_t mca_btl_openib_component = {
     {
         /* First, the mca_base_component_t struct containing meta information
@@ -507,71 +506,67 @@ int mca_btl_openib_component_progress()
          *   we will check the high priority and process them until there are none left. 
          *   note that low priority messages are only processed one per progress call. 
          */
-        do{
-            ne=ibv_poll_cq(openib_btl->ib_cq_high, 1, &wc );
-            if(ne < 0 ){ 
-                BTL_ERROR(("error polling CQ with %d errno says %s\n", ne, strerror(errno))); 
+        ne=ibv_poll_cq(openib_btl->ib_cq_high, 1, &wc );
+        if(ne < 0 ){ 
+            BTL_ERROR(("error polling CQ with %d errno says %s\n", ne, strerror(errno))); 
+            return OMPI_ERROR; 
+        } 
+        else if(wc.status != IBV_WC_SUCCESS) { 
+            BTL_ERROR(("error polling CQ with status %d for wr_id %llu\n", 
+                       wc.status, wc.wr_id)); 
+            return OMPI_ERROR; 
+        }
+        else if(1 == ne) { 
+            BTL_VERBOSE(("completion queue event says opcode is %d\n", wc.opcode)); 
+
+            /* Handle work completions */
+            switch(wc.opcode) {
+            case IBV_WC_RECV_RDMA_WITH_IMM: 
+                BTL_ERROR(("Got an RDMA with Immediate data Not supported!")); 
                 return OMPI_ERROR; 
-            } 
-            else if(wc.status != IBV_WC_SUCCESS) { 
-                BTL_ERROR(("error polling CQ with status %d for wr_id %llu\n", 
-                          wc.status, wc.wr_id)); 
-                return OMPI_ERROR; 
-            }
-            else if(1 == ne) { 
-                BTL_VERBOSE(("completion queue event says opcode is %d\n", wc.opcode)); 
+                
+            case IBV_WC_RECV: 
+                /* Process a RECV */ 
+                    
+                BTL_VERBOSE(("Got an  recv on the completion queue")); 
+                frag = (mca_btl_openib_frag_t*) wc.wr_id;
+                endpoint = (mca_btl_openib_endpoint_t*) frag->endpoint; 
+                frag->rc=OMPI_SUCCESS; 
+                frag->segment.seg_len =  
+                    wc.byte_len-
+                    ((unsigned char*) frag->segment.seg_addr.pval  - (unsigned char*) frag->hdr); 
+                
+                
+                    
+                /* advance the segment address past the header and subtract from the length..*/ 
+                openib_btl->ib_reg[frag->hdr->tag].cbfunc(&openib_btl->super, 
+                                                          frag->hdr->tag, 
+                                                          &frag->base, 
+                                                          openib_btl->ib_reg[frag->hdr->tag].cbdata);         
+                
+                OPAL_THREAD_ADD32(&endpoint->rr_posted_high, -1); 
+                MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0); 
+                OMPI_FREE_LIST_RETURN(&(openib_btl->recv_free_eager), (opal_list_item_t*) frag); 
+                count++; 
+                break; 
 
-                /* Handle work completions */
-                switch(wc.opcode) {
-                case IBV_WC_RECV_RDMA_WITH_IMM: 
-                    BTL_ERROR(("Got an RDMA with Immediate data Not supported!")); 
-                    return OMPI_ERROR; 
-                
-                case IBV_WC_RECV: 
-                    /* Process a RECV */ 
+            case IBV_WC_RDMA_READ: 
+            case IBV_WC_RDMA_WRITE: 
+            case IBV_WC_SEND :
                     
-                    BTL_VERBOSE(("Got an  recv on the completion queue")); 
-                    frag = (mca_btl_openib_frag_t*) wc.wr_id;
-                    endpoint = (mca_btl_openib_endpoint_t*) frag->endpoint; 
-                    frag->rc=OMPI_SUCCESS; 
-                    frag->segment.seg_len =  
-                        wc.byte_len-
-                        ((unsigned char*) frag->segment.seg_addr.pval  - (unsigned char*) frag->hdr); 
-                
-                
+                /* Process a completed send or rdma write*/
+                frag = (mca_btl_openib_frag_t*) wc.wr_id; 
+                frag->rc = OMPI_SUCCESS; 
+                frag->base.des_cbfunc(&openib_btl->super, frag->endpoint, &frag->base, frag->rc); 
+                count++;
+                break; 
                     
-                    /* advance the segment address past the header and subtract from the length..*/ 
-                    openib_btl->ib_reg[frag->hdr->tag].cbfunc(&openib_btl->super, 
-                                                              frag->hdr->tag, 
-                                                              &frag->base, 
-                                                              openib_btl->ib_reg[frag->hdr->tag].cbdata);         
-                
-                    OPAL_THREAD_ADD32(&endpoint->rr_posted_high, -1); 
-                    MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0); 
-                    OMPI_FREE_LIST_RETURN(&(openib_btl->recv_free_eager), (opal_list_item_t*) frag); 
-                    count++; 
-                    break; 
-
-                case IBV_WC_RDMA_WRITE: 
-                case IBV_WC_SEND :
-                    
-                    /* Process a completed send or rdma write*/
-                    frag = (mca_btl_openib_frag_t*) wc.wr_id; 
-                    frag->rc = OMPI_SUCCESS; 
-                    frag->base.des_cbfunc(&openib_btl->super, frag->endpoint, &frag->base, frag->rc); 
-                    count++;
-                    break; 
-                    
-
-                    break;
-                
-                default:
-                    BTL_ERROR(("Unhandled work completion opcode is %d", wc.opcode));
-                    break;
-                }
+                                    
+            default:
+                BTL_ERROR(("Unhandled work completion opcode is %d", wc.opcode));
+                break;
             }
         }
-        while(ne > 0); 
         
         ne=ibv_poll_cq(openib_btl->ib_cq_low, 1, &wc );
         if(ne < 0){ 
@@ -614,8 +609,10 @@ int mca_btl_openib_component_progress()
                 count++; 
                 break; 
 
+            case IBV_WC_RDMA_READ: 
             case IBV_WC_RDMA_WRITE: 
             case IBV_WC_SEND :
+
                 /* Process a completed send */
                 frag = (mca_btl_openib_frag_t*) wc.wr_id; 
                 frag->rc = OMPI_SUCCESS; 

ompi/mca/btl/openib/btl_openib_endpoint.c

@@ -32,7 +32,6 @@
 #include "class/ompi_free_list.h" 
 #include <errno.h> 
 #include <string.h> 
-extern int errno; 
 
 static void mca_btl_openib_endpoint_construct(mca_btl_base_endpoint_t* endpoint);
 static void mca_btl_openib_endpoint_destruct(mca_btl_base_endpoint_t* endpoint);
@@ -728,7 +727,7 @@ int mca_btl_openib_endpoint_create_qp(
         qp_attr->qp_state = IBV_QPS_INIT; 
         qp_attr->pkey_index = mca_btl_openib_component.ib_pkey_ix; 
         qp_attr->port_num = openib_btl->port_num; 
-        qp_attr->qp_access_flags = IBV_ACCESS_REMOTE_WRITE; 
+        qp_attr->qp_access_flags = IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ; 
         
         if(ibv_modify_qp((*qp), qp_attr, 
                          IBV_QP_STATE | 

ompi/mca/btl/openib/btl_openib_endpoint.h

@@ -27,7 +27,6 @@
 #include <errno.h> 
 #include <string.h> 
 #include "mca/btl/base/btl_base_error.h"
-extern int errno; 
 
 #if defined(c_plusplus) || defined(__cplusplus)
 extern "C" {

ompi/mca/mpool/openib/mpool_openib_component.c

@@ -87,16 +87,6 @@ OBJ_CLASS_INSTANCE(
 
 
 
-static char* mca_mpool_openib_param_register_string(
-    const char* param_name,
-    const char* default_value)
-{
-    char *param_value;
-    int id = mca_base_param_register_string("mpool","openib",param_name,NULL,default_value);
-    mca_base_param_lookup_string(id, &param_value);
-    return param_value;
-}
-
 
 /**
   * component open/close/init function

ompi/mca/mpool/openib/mpool_openib_module.c

@@ -46,10 +46,7 @@ void* mca_mpool_openib_alloc(
     size_t align, 
     mca_mpool_base_registration_t** registration)
 {
-    mca_mpool_openib_module_t* mpool_openib = (mca_mpool_openib_module_t*)mpool; 
-    /* void* addr_malloc = (void*)malloc((*size) + mca_mpool_openib_component.page_size);  */
-    /* void* addr = (void*)  ALIGN_ADDR(addr_malloc, mca_mpool_openib_component.page_size_log);  */
-   
+       
     void* addr_malloc = (void*)memalign(mca_mpool_openib_component.page_size, size); 
     void* addr = addr_malloc; 
 
@@ -81,7 +78,7 @@ int mca_mpool_openib_register(mca_mpool_base_module_t* mpool,
                               mpool_module->resources.ib_pd, 
                               addr, 
                               size, 
-                              IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE
+                              IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ
                               /* IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE */ 
                               ); 
    
@@ -104,7 +101,6 @@ int mca_mpool_openib_register(mca_mpool_base_module_t* mpool,
 int mca_mpool_openib_deregister(mca_mpool_base_module_t* mpool, void *addr, size_t size, 
                               mca_mpool_base_registration_t* registration){
     
-    mca_mpool_openib_module_t * mpool_openib = (mca_mpool_openib_module_t*) mpool; 
     mca_mpool_openib_registration_t * openib_reg; 
     openib_reg = (mca_mpool_openib_registration_t*) registration; 
     if(ibv_dereg_mr(openib_reg->mr)){   
@@ -127,7 +123,7 @@ void* mca_mpool_openib_realloc(
     mca_mpool_base_registration_t* old_reg  = *registration; 
     void* new_mem = mpool->mpool_alloc(mpool, size, 0, registration); 
     memcpy(new_mem, addr, old_reg->bound - old_reg->base); 
-    mpool->mpool_free(mpool, addr, &old_reg); 
+    mpool->mpool_free(mpool, addr, old_reg); 
     return new_mem; 
 
 }