If SM BTL cannot send fragment because the cyclic buffer is full put the

fragment on the pending list and send it later instead of spinning on
opal_progress().

This commit was SVN r16537.

ompi/mca/btl/sm/btl_sm.c

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2006-2007 Voltaire. All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -202,6 +203,7 @@ int mca_btl_sm_add_procs(
         peer->peer_smp_rank = n_local_procs +
             mca_btl_sm_component.num_smp_procs;
 
+        OBJ_CONSTRUCT(&peer->pending_sends, opal_list_t);
 #if OMPI_ENABLE_PROGRESS_THREADS == 1
         sprintf(path, "%s"OPAL_PATH_SEP"sm_fifo.%lu", orte_process_info.job_session_dir, 
                 (unsigned long)procs[proc]->proc_name.vpid);
@@ -575,6 +577,11 @@ int mca_btl_sm_add_procs(
                 mca_btl_sm_component.sm_free_list_inc,
                 NULL);
 
+        opal_free_list_init(&mca_btl_sm_component.pending_send_fl,
+                sizeof(btl_sm_pending_send_item_t),
+                OBJ_CLASS(opal_free_list_item_t),
+                16, -1, 32);
+
         /* set up mca_btl_sm_component.list_smp_procs */
         mca_btl_sm_component.list_smp_procs=(int *)
             malloc(mca_btl_sm_component.sm_max_procs*sizeof(int));
@@ -799,17 +806,21 @@ int mca_btl_sm_send(
     mca_btl_base_tag_t tag)
 {
     mca_btl_sm_frag_t* frag = (mca_btl_sm_frag_t*)descriptor;
+    int rc;
 
-    /* availble header space */
+    /* available header space */
     frag->hdr->len = frag->segment.seg_len;
     /* type of message, pt-2-pt, one-sided, etc */
     frag->hdr->tag = tag;
 
+    frag->endpoint = endpoint;
+
     /* 
      * post the descriptor in the queue - post with the relative
      * address 
      */
-    MCA_BTL_SM_FIFO_WRITE(endpoint, endpoint->my_smp_rank, endpoint->peer_smp_rank, frag->hdr);
+    MCA_BTL_SM_FIFO_WRITE(endpoint, endpoint->my_smp_rank,
+            endpoint->peer_smp_rank, frag->hdr, false, rc);
     return OMPI_SUCCESS;
 }
 

ompi/mca/btl/sm/btl_sm.h

@@ -10,6 +10,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2006-2007 Voltaire. All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -35,6 +36,7 @@
 #ifdef HAVE_UNISTD_H
 #include <unistd.h>
 #endif  /* HAVE_UNISTD_H */
+#include "opal/class/opal_free_list.h"
 #include "ompi/class/ompi_free_list.h"
 #include "ompi/class/ompi_bitmap.h"
 #include "ompi/class/ompi_fifo.h"
@@ -124,6 +126,8 @@ struct mca_btl_sm_component_t {
                                                     awaiting resources */
     struct mca_btl_base_endpoint_t **sm_peers;
 
+    opal_free_list_t pending_send_fl;
+
 #if OMPI_ENABLE_PROGRESS_THREADS == 1
     char sm_fifo_path[PATH_MAX];   /**< path to fifo used to signal this process */
     int  sm_fifo_fd;               /**< file descriptor corresponding to opened fifo */
@@ -133,6 +137,13 @@ struct mca_btl_sm_component_t {
 typedef struct mca_btl_sm_component_t mca_btl_sm_component_t;
 OMPI_MODULE_DECLSPEC extern mca_btl_sm_component_t mca_btl_sm_component;
 
+struct btl_sm_pending_send_item_t
+{
+    opal_free_list_item_t super;
+    void *data;
+};
+typedef struct btl_sm_pending_send_item_t btl_sm_pending_send_item_t;
+
 /**
  * Register shared memory module parameters with the MCA framework
  */

ompi/mca/btl/sm/btl_sm_component.c

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2006-2007 Voltaire. All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -178,6 +179,7 @@ int mca_btl_sm_component_open(void)
     OBJ_CONSTRUCT(&mca_btl_sm_component.sm_frags, ompi_free_list_t);
     OBJ_CONSTRUCT(&mca_btl_sm_component.sm_frags1, ompi_free_list_t);
     OBJ_CONSTRUCT(&mca_btl_sm_component.sm_frags2, ompi_free_list_t);
+    OBJ_CONSTRUCT(&mca_btl_sm_component.pending_send_fl, opal_free_list_t);
     return OMPI_SUCCESS;
 }
 
@@ -322,6 +324,48 @@ void mca_btl_sm_component_event_thread(opal_object_t* thread)
 }
 #endif
 
+void
+btl_sm_add_pending(struct mca_btl_base_endpoint_t *ep, void *data, bool resend)
+{
+    int rc;
+    btl_sm_pending_send_item_t *si;
+    opal_free_list_item_t *i;
+    OPAL_FREE_LIST_GET(&mca_btl_sm_component.pending_send_fl, i, rc);
+
+    /* don't handle error for now */
+    assert(i != NULL && rc == OMPI_SUCCESS);
+
+    si = (btl_sm_pending_send_item_t*)i;
+    si->data = data;
+
+    /* if data was on pending send list then prepend it to the list to
+     * minimize reordering */
+    if(resend)
+        opal_list_prepend(&ep->pending_sends, (opal_list_item_t*)si);
+    else
+        opal_list_append(&ep->pending_sends, (opal_list_item_t*)si);
+}
+
+static int process_pending_send(struct mca_btl_base_endpoint_t *ep)
+{
+    btl_sm_pending_send_item_t *si;
+    void *data;
+    opal_list_item_t *i = opal_list_remove_first(&ep->pending_sends);
+    int rc;
+
+    if(NULL == i) return OMPI_ERROR;
+
+    si = (btl_sm_pending_send_item_t*)i;
+
+    data = si->data;
+    OPAL_FREE_LIST_RETURN(&mca_btl_sm_component.pending_send_fl, i);
+
+    MCA_BTL_SM_FIFO_WRITE(ep, ep->my_smp_rank, ep->peer_smp_rank, data,
+            true, rc);
+
+    return rc;
+}
+
 int mca_btl_sm_component_progress(void)
 {
     /* local variables */
@@ -334,12 +378,6 @@ int mca_btl_sm_component_progress(void)
     int proc;
     int rc = 0;
 
-    /* send progress is made by the PML */
-
-    /* 
-     * receive progress 
-     */
-
     /* poll each fifo */
     for(proc = 0; proc < mca_btl_sm_component.num_smp_procs - 1; proc++) {
         peer_smp_rank = mca_btl_sm_component.list_smp_procs[proc];
@@ -377,6 +415,8 @@ int mca_btl_sm_component_progress(void)
                 /* completion callback */
                 frag->base.des_cbfunc(&mca_btl_sm.super, frag->endpoint,
                         &frag->base, status?OMPI_ERROR:OMPI_SUCCESS);
+                if(opal_list_get_size(&frag->endpoint->pending_sends))
+                    process_pending_send(frag->endpoint);
                 break;
             }
             case MCA_BTL_SM_FRAG_SEND:
@@ -397,7 +437,7 @@ int mca_btl_sm_component_progress(void)
                         reg->cbdata);
                 MCA_BTL_SM_FIFO_WRITE(
                         mca_btl_sm_component.sm_peers[peer_smp_rank],
-                        my_smp_rank, peer_smp_rank, hdr->frag);
+                        my_smp_rank, peer_smp_rank, hdr->frag, false, rc);
                 break;
             }
             default:
@@ -406,7 +446,7 @@ int mca_btl_sm_component_progress(void)
                         MCA_BTL_SM_FRAG_STATUS_MASK);
                 MCA_BTL_SM_FIFO_WRITE(
                         mca_btl_sm_component.sm_peers[peer_smp_rank],
-                        my_smp_rank, peer_smp_rank, hdr);
+                        my_smp_rank, peer_smp_rank, hdr, false, rc);
                 break;
         }
         rc++;

ompi/mca/btl/sm/btl_sm_endpoint.h

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2006-2007 Voltaire. All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -39,7 +40,9 @@ struct mca_btl_base_endpoint_t {
 #if OMPI_ENABLE_PROGRESS_THREADS == 1
     int fifo_fd;        /**< pipe/fifo used to signal endpoint that data is queued */
 #endif
+    opal_list_t pending_sends; /**< pending data to send */
 };
 
+void 
+btl_sm_add_pending(struct mca_btl_base_endpoint_t *ep, void *data, bool resend);
 #endif
-

ompi/mca/btl/sm/btl_sm_fifo.h

@@ -4,7 +4,8 @@
 #include "btl_sm.h"
 #include "btl_sm_endpoint.h"
 
-#define MCA_BTL_SM_FIFO_WRITE(endpoint_peer, my_smp_rank,peer_smp_rank,hdr) \
+#define MCA_BTL_SM_FIFO_WRITE(endpoint_peer, my_smp_rank,                   \
+        peer_smp_rank, hdr, resend, rc)                                     \
 do {                                                                        \
     ompi_fifo_t* fifo;                                                      \
     fifo=&(mca_btl_sm_component.fifo[peer_smp_rank][my_smp_rank]);          \
@@ -13,10 +14,14 @@ do {                                                                        \
     if(opal_using_threads())                                                \
         opal_atomic_lock(fifo->head_lock);                                  \
     /* post fragment */                                                     \
-    while(ompi_fifo_write_to_head(hdr, fifo,                                \
+    if(ompi_fifo_write_to_head(hdr, fifo, mca_btl_sm_component.sm_mpool)    \
-        mca_btl_sm_component.sm_mpool) != OMPI_SUCCESS)                     \
+            != OMPI_SUCCESS) {                                              \
-        opal_progress();                                                    \
+        btl_sm_add_pending(endpoint_peer, hdr, resend);                     \
+        rc = OMPI_ERR_TEMP_OUT_OF_RESOURCE;                                 \
+    } else {                                                                \
         MCA_BTL_SM_SIGNAL_PEER(endpoint_peer);                              \
+        rc = OMPI_SUCCESS;                                                  \
+    }                                                                       \
     if(opal_using_threads())                                                \
         opal_atomic_unlock(fifo->head_lock);                                \
 } while(0)