static inline int mca_btl_ugni_progress_rdma (mca_btl_ugni_module_t *ugni_module, mca_btl_ugni_device_t *device, mca_btl_ugni_cq_t *cq) { mca_btl_ugni_post_descriptor_t *post_desc[MCA_BTL_UGNI_COMPLETIONS_PER_LOOP]; gni_cq_entry_t event_data[MCA_BTL_UGNI_COMPLETIONS_PER_LOOP]; int rc; rc = mca_btl_ugni_cq_get_completed_desc (device, cq, event_data, post_desc, MCA_BTL_UGNI_COMPLETIONS_PER_LOOP); if (0 >= rc) { return rc; } BTL_VERBOSE(("got %d completed rdma descriptors", rc)); for (int i = 0 ; i < rc ; ++i) { BTL_VERBOSE(("post descriptor %p complete. GNI_CQ_STATUS_OK(): %d", post_desc[i], GNI_CQ_STATUS_OK(event_data[i]))); if (OPAL_UNLIKELY(!GNI_CQ_STATUS_OK(event_data[i]))) { uint32_t recoverable = 1; (void) GNI_CqErrorRecoverable (event_data[i], &recoverable); if (OPAL_UNLIKELY(++post_desc[i]->tries >= mca_btl_ugni_component.rdma_max_retries || !recoverable)) { char char_buffer[1024]; GNI_CqErrorStr (event_data[i], char_buffer, 1024); /* give up */ BTL_ERROR(("giving up on desciptor %p, recoverable %d: %s", (void *) post_desc[i], recoverable, char_buffer)); #if OPAL_ENABLE_DEBUG btl_ugni_dump_post_desc (post_desc[i]); #endif mca_btl_ugni_post_desc_complete (ugni_module, post_desc[i], OPAL_ERROR); return OPAL_ERROR; } mca_btl_ugni_repost (ugni_module, post_desc[i]); return 0; } mca_btl_ugni_post_desc_complete (ugni_module, post_desc[i], OPAL_SUCCESS); } /* should be resources to progress the pending post list */ (void) mca_btl_ugni_post_pending (ugni_module, device); return rc; }
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; }
int mca_btl_ugni_progress_remote_smsg (mca_btl_ugni_module_t *btl) { mca_btl_base_endpoint_t *ep; gni_cq_entry_t event_data; gni_return_t grc; uint64_t inst_id; grc = mca_btl_ugni_gni_cq_get_event (btl->devices, btl->smsg_remote_cq, &event_data); if (GNI_RC_NOT_DONE == grc) { return 0; } if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc || !GNI_CQ_STATUS_OK(event_data) || GNI_CQ_OVERRUN(event_data))) { if (GNI_RC_ERROR_RESOURCE == grc || (GNI_RC_SUCCESS == grc && GNI_CQ_OVERRUN(event_data))) { /* recover from smsg cq overrun */ return mca_btl_ugni_handle_remote_smsg_overrun (btl); } BTL_ERROR(("unhandled error in GNI_CqGetEvent")); /* unhandled error: crash */ assert (0); return mca_btl_rc_ugni_to_opal (grc); } BTL_VERBOSE(("REMOTE CQ: Got event 0x%" PRIx64 ". msg id = %" PRIu64 ". ok = %d, type = %" PRIu64, (uint64_t) event_data, GNI_CQ_GET_INST_ID(event_data), GNI_CQ_STATUS_OK(event_data), GNI_CQ_GET_TYPE(event_data))); inst_id = GNI_CQ_GET_INST_ID(event_data); ep = (mca_btl_base_endpoint_t *) opal_pointer_array_get_item (&btl->endpoints, inst_id); if (OPAL_UNLIKELY(MCA_BTL_UGNI_EP_STATE_CONNECTED != ep->state)) { /* due to the nature of datagrams we may get a smsg completion before we get mailbox info from the peer */ BTL_VERBOSE(("event occurred on an unconnected endpoint! ep state = %d", ep->state)); return 0; } return mca_btl_ugni_smsg_process (ep); }
static inline int mca_btl_ugni_progress_rdma (mca_btl_ugni_module_t *ugni_module) { ompi_common_ugni_post_desc_t *desc; mca_btl_ugni_base_frag_t *frag; gni_cq_entry_t event_data = 0; uint32_t recoverable = 1; gni_return_t rc; rc = GNI_CqGetEvent (ugni_module->rdma_local_cq, &event_data); if (GNI_RC_NOT_DONE == rc) { return 0; } if (OPAL_UNLIKELY((GNI_RC_SUCCESS != rc && !event_data) || GNI_CQ_OVERRUN(event_data))) { /* TODO -- need to handle overrun -- how do we do this without an event? will the event eventually come back? Ask Cray */ BTL_ERROR(("unhandled post error! ugni rc = %d", rc)); assert (0); return ompi_common_rc_ugni_to_ompi (rc); } rc = GNI_GetCompleted (ugni_module->rdma_local_cq, event_data, (gni_post_descriptor_t **) &desc); if (OPAL_UNLIKELY(GNI_RC_SUCCESS != rc && GNI_RC_TRANSACTION_ERROR != rc)) { BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[rc])); return ompi_common_rc_ugni_to_ompi (rc); } frag = MCA_BTL_UGNI_DESC_TO_FRAG(desc); if (OPAL_UNLIKELY(GNI_RC_SUCCESS != rc || !GNI_CQ_STATUS_OK(event_data))) { (void) GNI_CqErrorRecoverable (event_data, &recoverable); if (OPAL_UNLIKELY(++frag->post_desc.tries >= mca_btl_ugni_component.rdma_max_retries || !recoverable)) { /* give up */ BTL_ERROR(("giving up on frag %p", (void *) frag)); frag->cbfunc (frag, OMPI_ERROR); return OMPI_ERROR; } /* repost transaction */ mca_btl_ugni_repost (frag, OMPI_SUCCESS); return 0; } BTL_VERBOSE(("RDMA/FMA complete for frag %p", (void *) frag)); frag->cbfunc (frag, ompi_common_rc_ugni_to_ompi (rc)); return 1; }
static inline int mca_btl_ugni_progress_rdma (mca_btl_ugni_module_t *ugni_module, int which_cq) { mca_btl_ugni_post_descriptor_t *post_desc = NULL; gni_cq_entry_t event_data = 0; gni_post_descriptor_t *desc; uint32_t recoverable = 1; gni_return_t grc; gni_cq_handle_t the_cq; the_cq = (which_cq == 0) ? ugni_module->rdma_local_cq : ugni_module->rdma_local_irq_cq; OPAL_THREAD_LOCK(&ugni_module->device->dev_lock); grc = GNI_CqGetEvent (the_cq, &event_data); if (GNI_RC_NOT_DONE == grc) { OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock); return 0; } if (OPAL_UNLIKELY((GNI_RC_SUCCESS != grc && !event_data) || GNI_CQ_OVERRUN(event_data))) { /* TODO -- need to handle overrun -- how do we do this without an event? will the event eventually come back? Ask Cray */ BTL_ERROR(("unhandled post error! ugni rc = %d %s", grc, gni_err_str[grc])); OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock); return opal_common_rc_ugni_to_opal (grc); } grc = GNI_GetCompleted (the_cq, event_data, &desc); OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock); if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc && GNI_RC_TRANSACTION_ERROR != grc)) { BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[grc])); return opal_common_rc_ugni_to_opal (grc); } post_desc = MCA_BTL_UGNI_DESC_TO_PDESC(desc); if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc || !GNI_CQ_STATUS_OK(event_data))) { (void) GNI_CqErrorRecoverable (event_data, &recoverable); if (OPAL_UNLIKELY(++post_desc->desc.tries >= mca_btl_ugni_component.rdma_max_retries || !recoverable)) { char char_buffer[1024]; GNI_CqErrorStr (event_data, char_buffer, 1024); /* give up */ BTL_ERROR(("giving up on desciptor %p, recoverable %d: %s", (void *) post_desc, recoverable, char_buffer)); #if OPAL_ENABLE_DEBUG btl_ugni_dump_post_desc (post_desc); #endif mca_btl_ugni_post_desc_complete (ugni_module, post_desc, OPAL_ERROR); return OPAL_ERROR; } mca_btl_ugni_repost (ugni_module, post_desc); return 0; } mca_btl_ugni_post_desc_complete (ugni_module, post_desc, opal_common_rc_ugni_to_opal (grc)); return 1; }