ucs_status_t progress_remote_cq(uct_ugni_smsg_iface_t *iface) { gni_return_t ugni_rc; gni_cq_entry_t event_data; uct_ugni_ep_t tl_ep; uct_ugni_ep_t *ugni_ep; uct_ugni_smsg_ep_t *ep; ugni_rc = GNI_CqGetEvent(iface->remote_cq, &event_data); if(GNI_RC_NOT_DONE == ugni_rc){ return UCS_OK; } if (GNI_RC_SUCCESS != ugni_rc || !GNI_CQ_STATUS_OK(event_data) || GNI_CQ_OVERRUN(event_data)) { if(GNI_RC_ERROR_RESOURCE == ugni_rc || (GNI_RC_SUCCESS == ugni_rc && GNI_CQ_OVERRUN(event_data))){ ucs_debug("Detected remote CQ overrun. ungi_rc = %d [%s]", ugni_rc, gni_err_str[ugni_rc]); uct_ugni_smsg_handle_remote_overflow(iface); return UCS_OK; } ucs_error("GNI_CqGetEvent falied with unhandled error. Error status %s %d ", gni_err_str[ugni_rc], ugni_rc); return UCS_ERR_IO_ERROR; } tl_ep.hash_key = GNI_CQ_GET_INST_ID(event_data); ugni_ep = sglib_hashed_uct_ugni_ep_t_find_member(iface->super.eps, &tl_ep); ep = ucs_derived_of(ugni_ep, uct_ugni_smsg_ep_t); process_mbox(iface, ep); return UCS_INPROGRESS; }
static void uct_ugni_smsg_handle_remote_overflow(uct_ugni_smsg_iface_t *iface){ gni_return_t ugni_rc; gni_cq_entry_t event_data; struct sglib_hashed_uct_ugni_ep_t_iterator ep_iterator; uct_ugni_ep_t *current_ep; uct_ugni_smsg_ep_t *ep; /* We don't know which EP dropped a completion entry, so flush everything */ do{ ugni_rc = GNI_CqGetEvent(iface->remote_cq, &event_data); } while(GNI_RC_NOT_DONE != ugni_rc); current_ep = sglib_hashed_uct_ugni_ep_t_it_init(&ep_iterator, iface->super.eps); while(NULL != current_ep){ ep = ucs_derived_of(current_ep, uct_ugni_smsg_ep_t); process_mbox(iface, ep); current_ep = sglib_hashed_uct_ugni_ep_t_it_next(&ep_iterator); } }