portals4: use a separate EQ per PT

Signed-off-by: Antonio Pena Monferrer <apenya@mcs.anl.gov>

src/mpid/ch3/channels/nemesis/netmod/portals4/ptl_impl.h

@@ -19,8 +19,12 @@ extern ptl_pt_index_t  MPIDI_nem_ptl_pt;
 extern ptl_pt_index_t  MPIDI_nem_ptl_get_pt; /* portal for gets by receiver */
 extern ptl_pt_index_t  MPIDI_nem_ptl_control_pt; /* portal for MPICH control messages */
 extern ptl_pt_index_t  MPIDI_nem_ptl_rpt_pt; /* portal for MPICH control messages */
-extern ptl_handle_eq_t MPIDI_nem_ptl_target_eq;
 extern ptl_handle_eq_t MPIDI_nem_ptl_origin_eq;
+extern ptl_handle_eq_t MPIDI_nem_ptl_eq;
+extern ptl_handle_eq_t MPIDI_nem_ptl_get_eq;
+extern ptl_handle_eq_t MPIDI_nem_ptl_control_eq;
+extern ptl_handle_eq_t MPIDI_nem_ptl_origin_eq;
+extern ptl_handle_eq_t MPIDI_nem_ptl_rpt_eq;
 
 extern ptl_handle_md_t MPIDI_nem_ptl_global_md;
 extern ptl_ni_limits_t MPIDI_nem_ptl_ni_limits;

src/mpid/ch3/channels/nemesis/netmod/portals4/ptl_init.c

@@ -28,7 +28,9 @@ ptl_pt_index_t  MPIDI_nem_ptl_pt;
 ptl_pt_index_t  MPIDI_nem_ptl_get_pt; /* portal for gets by receiver */
 ptl_pt_index_t  MPIDI_nem_ptl_control_pt; /* portal for MPICH control messages */
 ptl_pt_index_t  MPIDI_nem_ptl_rpt_pt; /* portal for rportals control messages */
-ptl_handle_eq_t MPIDI_nem_ptl_target_eq;
+ptl_handle_eq_t MPIDI_nem_ptl_eq;
+ptl_handle_eq_t MPIDI_nem_ptl_get_eq;
+ptl_handle_eq_t MPIDI_nem_ptl_control_eq;
 ptl_handle_eq_t MPIDI_nem_ptl_origin_eq;
 ptl_pt_index_t  MPIDI_nem_ptl_control_rpt_pt; /* portal for rportals control messages */
 ptl_pt_index_t  MPIDI_nem_ptl_get_rpt_pt; /* portal for rportals control messages */
@@ -184,7 +186,14 @@ static int ptl_init(MPIDI_PG_t *pg_p, int pg_rank, char **bc_val_p, int *val_max
                     PTL_PID_ANY, &desired, &MPIDI_nem_ptl_ni_limits, &MPIDI_nem_ptl_ni);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
 
-    ret = PtlEQAlloc(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_ni_limits.max_eqs, &MPIDI_nem_ptl_target_eq);
+    /* allocate EQs for each portal */
+    ret = PtlEQAlloc(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_ni_limits.max_eqs, &MPIDI_nem_ptl_eq);
+    MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
+
+    ret = PtlEQAlloc(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_ni_limits.max_eqs, &MPIDI_nem_ptl_get_eq);
+    MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
+
+    ret = PtlEQAlloc(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_ni_limits.max_eqs, &MPIDI_nem_ptl_control_eq);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate a separate EQ for origin events. with this, we can implement rate-limit operations
@@ -193,32 +202,32 @@ static int ptl_init(MPIDI_PG_t *pg_p, int pg_rank, char **bc_val_p, int *val_max
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for matching messages */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for large messages where receiver does a get */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_get_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_get_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for MPICH control messages */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_control_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_control_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for MPICH control messages */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_rpt_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for MPICH control messages */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_get_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_get_rpt_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
     /* allocate portal for MPICH control messages */
-    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_target_eq,
+    ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_control_eq,
                      PTL_PT_ANY, &MPIDI_nem_ptl_control_rpt_pt);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
 
@@ -276,7 +285,7 @@ static int ptl_finalize(void)
 {
     int mpi_errno = MPI_SUCCESS;
     int ret;
-    ptl_handle_eq_t eqs[2];
+    ptl_handle_eq_t eqs[4];
     MPIDI_STATE_DECL(MPID_STATE_PTL_FINALIZE);
     MPIDI_FUNC_ENTER(MPID_STATE_PTL_FINALIZE);
 
@@ -288,9 +297,11 @@ static int ptl_finalize(void)
     if (mpi_errno) MPIU_ERR_POP(mpi_errno);
 
     /* shut down portals */
-    eqs[0] = MPIDI_nem_ptl_target_eq;
-    eqs[1] = MPIDI_nem_ptl_origin_eq;
-    ret = MPID_nem_ptl_rptl_drain_eq(2, eqs);
+    eqs[0] = MPIDI_nem_ptl_eq;
+    eqs[1] = MPIDI_nem_ptl_get_eq;
+    eqs[2] = MPIDI_nem_ptl_control_eq;
+    eqs[3] = MPIDI_nem_ptl_origin_eq;
+    ret = MPID_nem_ptl_rptl_drain_eq(4, eqs);
     MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptfree", "**ptlptfree %s", MPID_nem_ptl_strerror(ret));
 
     ret = MPID_nem_ptl_rptl_ptfini(MPIDI_nem_ptl_pt);

src/mpid/ch3/channels/nemesis/netmod/portals4/ptl_poll.c

@@ -131,17 +131,27 @@ int MPID_nem_ptl_poll(int is_blocking_poll)
     /* MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_POLL); */
 
     while (1) {
-        /* check both origin and target EQs for events */
-        ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_target_eq, &event);
+        /* check EQs for events */
+        ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_eq, &event);
         MPIU_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
+        if (ret == PTL_EQ_EMPTY) {
+            ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_get_eq, &event);
+            MPIU_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
+
+            if (ret == PTL_EQ_EMPTY) {
+                ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_control_eq, &event);
+                MPIU_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
+
                 if (ret == PTL_EQ_EMPTY) {
                     ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_origin_eq, &event);
                     MPIU_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
+                }
 
-            /* if both queues are empty, exit the loop */
+                /* all EQs are empty */
                 if (ret == PTL_EQ_EMPTY)
                     break;
             }
+        }
         MPIU_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqget", "**ptleqget %s", MPID_nem_ptl_strerror(ret));
         MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "Received event %s pt_idx=%d ni_fail=%s list=%s user_ptr=%p hdr_data=%#lx mlength=%lu rlength=%lu",
                                                 MPID_nem_ptl_strevent(&event), event.pt_index, MPID_nem_ptl_strnifail(event.ni_fail_type),