mtl/portals4: add timeout to get retransmit

Signed-off-by: Todd Kordenbrock <thkgcode@gmail.com>

Author: plesn

ompi/mca/mtl/portals4/mtl_portals4.h

@@ -73,6 +73,7 @@ struct mca_mtl_portals4_module_t {
 
     /* free list of rendezvous get fragments */
     opal_free_list_t fl_rndv_get_frag;
+    int get_retransmit_timeout;
 
     /** Network interface handle for matched interface */
     ptl_handle_ni_t ni_h;

ompi/mca/mtl/portals4/mtl_portals4_component.c

@@ -202,6 +202,16 @@ ompi_mtl_portals4_component_register(void)
                                            MCA_BASE_VAR_SCOPE_READONLY,
                                            &ompi_mtl_portals4.max_msg_size_mtl);
 
+    ompi_mtl_portals4.get_retransmit_timeout=10000;
+    (void) mca_base_component_var_register(&mca_mtl_portals4_component.mtl_version,
+                                           "get_retransmit_timeout",
+                                           "PtlGET retransmission timeout in usec",
+                                           MCA_BASE_VAR_TYPE_INT,
+                                           NULL, 0, 0,
+                                           OPAL_INFO_LVL_5,
+                                           MCA_BASE_VAR_SCOPE_READONLY,
+                                           &ompi_mtl_portals4.get_retransmit_timeout);
+
     OBJ_RELEASE(new_enum);
     if (0 > ret) {
         return OMPI_ERR_NOT_SUPPORTED;

ompi/mca/mtl/portals4/mtl_portals4_recv.c

@@ -27,6 +27,7 @@
 #include "ompi/mca/mtl/base/base.h"
 #include "ompi/mca/mtl/base/mtl_base_datatype.h"
 #include "ompi/message/message.h"
+#include "opal/mca/timer/base/base.h"
 
 #include "mtl_portals4.h"
 #include "mtl_portals4_endpoint.h"
@@ -81,6 +82,7 @@ read_msg(void *start, ptl_size_t length, ptl_process_t target,
         frag->frag_remote_offset = remote_offset + i * ompi_mtl_portals4.max_msg_size_mtl;
 
         frag->event_callback = ompi_mtl_portals4_rndv_get_frag_progress;
+        frag->frag_start_time_usec = opal_timer_base_get_usec();
 
         OPAL_OUTPUT_VERBOSE((90, ompi_mtl_base_framework.framework_output, "GET (fragment %d/%d, size %ld) send",
                              i + 1, frag_count, frag->frag_length));
@@ -322,17 +324,24 @@ ompi_mtl_portals4_rndv_get_frag_progress(ptl_event_t *ev,
     ompi_mtl_portals4_recv_request_t* ptl_request =
         (ompi_mtl_portals4_recv_request_t*) rndv_get_frag->request;
 
-    assert(ev->type==PTL_EVENT_REPLY);
+    assert(PTL_EVENT_REPLY == ev->type);
 
     OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
         "Recv %lu (0x%lx) got reply event",
         ptl_request->opcount, ptl_request->hdr_data));
 
+
     if (OPAL_UNLIKELY(ev->ni_fail_type != PTL_NI_OK)) {
         opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
                             "%s:%d: PTL_EVENT_REPLY with ni_fail_type: %d",
                             __FILE__, __LINE__, ev->ni_fail_type);
 
+        opal_timer_t time = opal_timer_base_get_usec() - rndv_get_frag->frag_start_time_usec;
+        if (time > (unsigned int) ompi_mtl_portals4.get_retransmit_timeout) {
+            mtl_ptl_error(1, "timeout retrying GET");
+            return OMPI_ERROR;
+        }
+
         OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
             "Rendezvous Get Failed: Reissuing frag #%u", rndv_get_frag->frag_num));
 

ompi/mca/mtl/portals4/mtl_portals4_request.h

@@ -22,6 +22,7 @@
 
 #include "opal/datatype/opal_convertor.h"
 #include "ompi/mca/mtl/mtl.h"
+#include "opal/mca/timer/base/base.h"
 
 struct ompi_mtl_portals4_message_t;
 struct ompi_mtl_portals4_pending_request_t;
@@ -93,6 +94,7 @@ struct ompi_mtl_portals4_rndv_get_frag_t {
     ptl_process_t    frag_target;
     ptl_hdr_data_t   frag_match_bits;
     ptl_size_t       frag_remote_offset;
+    opal_timer_t     frag_start_time_usec;
 
     int (*event_callback)(ptl_event_t *ev, struct ompi_mtl_portals4_rndv_get_frag_t*);