static void __osmv_dispatch_simple_mad(IN osm_bind_handle_t h_bind, IN const ib_mad_t * p_mad, IN osmv_txn_ctx_t * p_txn, IN const osm_mad_addr_t * p_mad_addr) { osm_madw_t *p_madw; ib_mad_t *p_mad_buf; osm_madw_t *p_req_madw = NULL; osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; OSM_LOG_ENTER(p_bo->p_vendor->p_log); /* Build the MAD wrapper to be returned to the user. * The actual storage for the MAD is allocated there. */ p_madw = osm_mad_pool_get(p_bo->p_osm_pool, h_bind, MAD_BLOCK_SIZE, p_mad_addr); if (NULL == p_madw) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_dispatch_simple_mad: ERR 6501: " "Out Of Memory - could not allocate a buffer of size %d\n", MAD_BLOCK_SIZE); goto dispatch_simple_mad_done; } p_mad_buf = osm_madw_get_mad_ptr(p_madw); /* Copy the payload to the MAD buffer */ memcpy((void *)p_mad_buf, (void *)p_mad, MAD_BLOCK_SIZE); if (NULL != p_txn) { /* This is a RESPONSE MAD. Pair it with the REQUEST MAD, pass upstream */ p_req_madw = p_txn->p_madw; CL_ASSERT(NULL != p_req_madw); p_mad_buf->trans_id = cl_hton64(osmv_txn_get_tid(p_txn)); osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "Restoring the original TID to 0x%" PRIx64 "\n", cl_ntoh64(p_mad_buf->trans_id)); /* Reply matched, transaction complete */ osmv_txn_done(h_bind, osmv_txn_get_key(p_txn), FALSE); } else { /* This is a REQUEST MAD. Don't create a context, pass upstream */ } /* Do the job ! */ p_bo->recv_cb(p_madw, p_bo->cb_context, p_req_madw); dispatch_simple_mad_done: OSM_LOG_EXIT(p_bo->p_vendor->p_log); }
void __osmv_txn_all_done(osm_bind_handle_t h_bind) { osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; cl_map_item_t *p_item; cl_map_obj_t *p_obj; osmv_txn_ctx_t *p_txn; OSM_LOG_ENTER(p_bo->p_vendor->p_log); p_item = cl_qmap_head(p_bo->txn_mgr.p_txn_map); while (p_item != cl_qmap_end(p_bo->txn_mgr.p_txn_map)) { p_obj = PARENT_STRUCT(p_item, cl_map_obj_t, item); p_txn = (osmv_txn_ctx_t *) cl_qmap_obj(p_obj); osmv_txn_done(h_bind, osmv_txn_get_key(p_txn), FALSE); free(p_obj); /* assuming osmv_txn_done has removed the txn from the map */ p_item = cl_qmap_head(p_bo->txn_mgr.p_txn_map); } OSM_LOG_EXIT(p_bo->p_vendor->p_log); }
static uint64_t __osmv_txn_timeout_cb(IN uint64_t key, IN uint32_t num_regs, IN void *cb_context) { osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) cb_context; uint64_t ret = 0; osmv_txn_ctx_t *p_txn; osmv_rmpp_send_ctx_t *p_send_ctx; osm_madw_t *p_madw = NULL; ib_mad_t *p_mad; osm_mad_addr_t *p_mad_addr; boolean_t invoke_err_cb = FALSE; OSM_LOG_ENTER(p_bo->p_vendor->p_log); /* Don't try to acquire a lock on the Bind Object - * it's taken by the mechanism that drives the timeout based events! * (Recall the special constructor that the Event Wheel is applied with) */ if (p_bo->is_closing) { goto txn_done; } ret = osmv_txn_lookup(p_bo, key, &p_txn); if (IB_NOT_FOUND == ret) { /* Prevent a race - the transaction is already destroyed */ goto txn_done; } p_madw = p_txn->p_madw; switch (osmv_txn_get_rmpp_state(p_txn)) { case OSMV_TXN_RMPP_NONE: if (num_regs <= OSMV_MAX_RETRANSMIT) { /* We still did not exceed the limit of retransmissions. * Set the next timeout's value. */ osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "__osmv_txn_timeout_cb: " "The transaction request (tid=0x%llX) timed out %d times. " "Retrying the send.\n", osmv_txn_get_tid(p_txn), num_regs); /* resend this mad */ ret = osmv_simple_send_madw((osm_bind_handle_t *) p_bo, p_madw, p_txn, TRUE); if (ret != IB_SUCCESS) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_txn_timeout_cb: " "Fail to send retry for transaction request (tid=0x%llX).\n", osmv_txn_get_tid(p_txn)); osmv_txn_done((osm_bind_handle_t) p_bo, key, TRUE /*in timeout callback */ ); /* This is a requester. Always apply the callback */ invoke_err_cb = TRUE; } else { uint64_t next_timeout_ms; next_timeout_ms = p_bo->p_vendor->resp_timeout * (num_regs + 1) * (num_regs + 1); /* when do we need to timeout again */ ret = cl_get_time_stamp() + (uint64_t) (1000 * next_timeout_ms); osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "__osmv_txn_timeout_cb: " "Retry request timout in : %lu [msec].\n", next_timeout_ms); } } else { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_txn_timeout_cb: ERR 6702: " "The transaction request (tid=0x%llX) timed out (after %d retries). " "Invoking the error callback.\n", osmv_txn_get_tid(p_txn), num_regs); osmv_txn_done((osm_bind_handle_t) p_bo, key, TRUE /*in timeout callback */ ); /* This is a requester. Always apply the callback */ invoke_err_cb = TRUE; } break; case OSMV_TXN_RMPP_SENDER: osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "RMPP sender (tid=0x%llX) did not receive ACK " "on every segment in the current send window.\n", osmv_txn_get_tid(p_txn)); p_send_ctx = osmv_txn_get_rmpp_send_ctx(p_txn); if (num_regs <= OSMV_MAX_RETRANSMIT) { /* We still did not exceed the limit of retransmissions. * Set the next timeout's value. */ ret = cl_get_time_stamp() + 1000 * p_bo->p_vendor->resp_timeout; } else { p_send_ctx->status = IB_TIMEOUT; p_mad = osm_madw_get_mad_ptr(p_madw); p_mad_addr = osm_madw_get_mad_addr_ptr(p_madw); /* Send an ABORT to the other side */ osmv_rmpp_send_nak((osm_bind_handle_t) p_bo, p_mad, p_mad_addr, IB_RMPP_TYPE_ABORT, IB_RMPP_STATUS_T2L); } /* Wake the RMPP sender thread up */ cl_event_signal(&p_send_ctx->event); break; case OSMV_TXN_RMPP_RECEIVER: osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "Transaction timeout on an RMPP receiver (tid=0x%llX). " "Dropping the transaction.\n", osmv_txn_get_tid(p_txn)); osmv_txn_done((osm_bind_handle_t) p_bo, key, TRUE /*in timeout callback */ ); if (FALSE == osmv_txn_is_rmpp_init_by_peer(p_txn)) { /* This is a requester, still waiting for the reply. Apply the callback */ invoke_err_cb = TRUE; } break; default: CL_ASSERT(FALSE); } if (TRUE == invoke_err_cb) { CL_ASSERT(NULL != p_madw); /* update the status in the p_madw */ p_madw->status = IB_TIMEOUT; p_bo->send_err_cb(p_bo->cb_context, p_madw); /* no re-registration */ ret = 0; } txn_done: OSM_LOG_EXIT(p_bo->p_vendor->p_log); return ret; }
static ib_api_status_t __osmv_get_send_txn(IN osm_bind_handle_t h_bind, IN osm_madw_t * const p_madw, IN boolean_t is_rmpp, IN boolean_t resp_expected, OUT osmv_txn_ctx_t ** pp_txn) { ib_api_status_t ret; uint64_t tid, key; osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; ib_mad_t *p_mad = osm_madw_get_mad_ptr(p_madw); OSM_LOG_ENTER(p_bo->p_vendor->p_log); CL_ASSERT(NULL != pp_txn); key = tid = cl_ntoh64(p_mad->trans_id); if (TRUE == resp_expected) { /* Create a unique identifier at the requester side */ key = osmv_txn_uniq_key(tid); } /* We must run under a transaction framework */ ret = osmv_txn_lookup(h_bind, key, pp_txn); if (IB_NOT_FOUND == ret) { /* Generally, we start a new transaction */ ret = osmv_txn_init(h_bind, tid, key, pp_txn); if (IB_SUCCESS != ret) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_get_send_txn: ERR 7313: " "The transaction id=0x%" PRIx64 " failed to init.\n", tid); goto get_send_txn_done; } } else { CL_ASSERT(NULL != *pp_txn); /* The transaction context exists. * This is legal only if I am going to return an * (RMPP?) reply to an RMPP request sent by the other part * (double-sided RMPP transfer) */ if (FALSE == is_rmpp || FALSE == osmv_txn_is_rmpp_init_by_peer(*pp_txn)) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_get_send_txn: ERR 7314: " "The transaction id=0x%" PRIx64 " is not unique. Send failed.\n", tid); ret = IB_INVALID_SETTING; goto get_send_txn_done; } if (TRUE == resp_expected) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_get_send_txn: ERR 7315: " "The transaction id=0x%" PRIx64 " can't expect a response. Send failed.\n", tid); ret = IB_INVALID_PARAMETER; goto get_send_txn_done; } } if (TRUE == is_rmpp) { ret = osmv_txn_init_rmpp_sender(h_bind, *pp_txn, p_madw); if (IB_SUCCESS != ret) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_get_send_txn: ERR 7316: " "The transaction id=0x%" PRIx64 " failed to init the rmpp mad. Send failed.\n", tid); osmv_txn_done(h_bind, tid, FALSE); goto get_send_txn_done; } } /* Save a reference to the MAD in the txn context * We'll need to match it in two cases: * (1) When the response is returned, if I am the requester * (2) In RMPP retransmissions */ osmv_txn_set_madw(*pp_txn, p_madw); get_send_txn_done: OSM_LOG_EXIT(p_bo->p_vendor->p_log); return ret; }
ib_api_status_t osm_vendor_send(IN osm_bind_handle_t h_bind, IN osm_madw_t * const p_madw, IN boolean_t const resp_expected) { ib_api_status_t ret = IB_SUCCESS; osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; boolean_t is_rmpp = FALSE, is_rmpp_ds = FALSE; osmv_txn_ctx_t *p_txn = NULL; ib_mad_t *p_mad; osm_log_t *p_log = p_bo->p_vendor->p_log; osm_mad_pool_t *p_mad_pool = p_bo->p_osm_pool; OSM_LOG_ENTER(p_log); if (NULL == h_bind || NULL == p_madw || NULL == (p_mad = osm_madw_get_mad_ptr(p_madw)) || NULL == osm_madw_get_mad_addr_ptr(p_madw)) { return IB_INVALID_PARAMETER; } is_rmpp = (p_madw->mad_size > MAD_BLOCK_SIZE || osmv_mad_is_rmpp(p_mad)); /* is this rmpp double sided? This means we expect a response that can be an rmpp or not */ is_rmpp_ds = (TRUE == is_rmpp && TRUE == resp_expected); /* Make our operations with the send context atomic */ osmv_txn_lock(p_bo); if (TRUE == p_bo->is_closing) { osm_log(p_log, OSM_LOG_ERROR, "osm_vendor_send: ERR 7310: " "The handle %p is being unbound, cannot send.\n", h_bind); ret = IB_INTERRUPTED; /* When closing p_bo could be detroyed or is going to , thus could not refer to it */ goto send_done; } if (TRUE == resp_expected || TRUE == is_rmpp) { /* We must run under a transaction framework. * Get the transaction object (old or new) */ ret = __osmv_get_send_txn(h_bind, p_madw, is_rmpp, resp_expected, &p_txn); if (IB_SUCCESS != ret) { goto send_done; } } if (TRUE == is_rmpp) { /* Do the job - RMPP! * The call returns as all the packets are ACK'ed/upon error * The txn lock will be released each time the function sleeps * and re-acquired when it wakes up */ ret = osmv_rmpp_send_madw(h_bind, p_madw, p_txn, is_rmpp_ds); } else { /* Do the job - single MAD! * The call returns as soon as the MAD is put on the wire */ ret = osmv_simple_send_madw(h_bind, p_madw, p_txn, FALSE); } if (IB_SUCCESS == ret) { if ((TRUE == is_rmpp) && (FALSE == is_rmpp_ds)) { /* For double-sided sends, the txn continues to live */ osmv_txn_done(h_bind, osmv_txn_get_key(p_txn), FALSE /*not in callback */ ); } if (FALSE == resp_expected) { osm_mad_pool_put(p_mad_pool, p_madw); } } else if (IB_INTERRUPTED != ret) { if (NULL != p_txn) { osmv_txn_done(h_bind, osmv_txn_get_key(p_txn), FALSE /*not in callback */ ); } osm_log(p_log, OSM_LOG_ERROR, "osm_vendor_send: ERR 7311: failed to send MADW %p\n", p_madw); if (TRUE == resp_expected) { /* Change the status on the p_madw */ p_madw->status = ret; /* Only the requester expects the error callback */ p_bo->send_err_cb(p_bo->cb_context, p_madw); } else { /* put back the mad - it is useless ... */ osm_mad_pool_put(p_mad_pool, p_madw); } } else { /* the transaction was aborted due to p_bo exit */ osm_mad_pool_put(p_mad_pool, p_madw); goto aborted; } send_done: osmv_txn_unlock(p_bo); aborted: OSM_LOG_EXIT(p_log); return ret; }
static ib_api_status_t __osmv_dispatch_rmpp_rcv(IN osm_bind_handle_t h_bind, IN const ib_mad_t * p_mad, IN osmv_txn_ctx_t * p_txn, IN const osm_mad_addr_t * p_mad_addr) { ib_api_status_t status = IB_SUCCESS; osmv_rmpp_recv_ctx_t *p_recv_ctx = osmv_txn_get_rmpp_recv_ctx(p_txn); osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; boolean_t is_last1 = FALSE, is_last2 = FALSE; osm_madw_t *p_new_madw = NULL, *p_req_madw = NULL; ib_mad_t *p_mad_buf; uint32_t size = 0; uint64_t key = osmv_txn_get_key(p_txn); uint64_t tid = osmv_txn_get_tid(p_txn); OSM_LOG_ENTER(p_bo->p_vendor->p_log); if (TRUE == osmv_rmpp_is_ack(p_mad)) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "Not supposed to receive ACK's --> dropping the MAD\n"); goto dispatch_rmpp_rcv_done; } if (TRUE == osmv_rmpp_is_abort_stop(p_mad)) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "__osmv_dispatch_rmpp_rcv: ERR 6504: " "The Remote Side stopped sending\n"); status = IB_REMOTE_ERROR; goto dispatch_rmpp_rcv_done; } status = __osmv_dispatch_accept_seg(h_bind, p_txn, p_mad); switch (status) { case IB_SUCCESS: /* Check wheter this is the legal last MAD */ /* Criteria #1: the received MAD is marked last */ is_last1 = osmv_rmpp_is_last(p_mad); /* Criteria #2: the total accumulated length hits the advertised one */ is_last2 = is_last1; size = osmv_rmpp_recv_ctx_get_byte_num_from_first(p_recv_ctx); if (size > 0) { is_last2 = (osmv_rmpp_recv_ctx_get_cur_byte_num(p_recv_ctx) >= size); } if (is_last1 != is_last2) { osmv_rmpp_send_nak(h_bind, p_mad, p_mad_addr, IB_RMPP_TYPE_ABORT, IB_RMPP_STATUS_BAD_LEN); status = IB_ERROR; goto dispatch_rmpp_rcv_done; } /* TBD Consider an optimization - sending an ACK * only for the last segment in the window */ __osmv_dispatch_send_ack(h_bind, p_mad, p_txn, p_mad_addr); break; case IB_INSUFFICIENT_RESOURCES: /* An out-of-order segment received. Send the ACK anyway */ __osmv_dispatch_send_ack(h_bind, p_mad, p_txn, p_mad_addr); status = IB_SUCCESS; goto dispatch_rmpp_rcv_done; case IB_INSUFFICIENT_MEMORY: osmv_rmpp_send_nak(h_bind, p_mad, p_mad_addr, IB_RMPP_TYPE_STOP, IB_RMPP_STATUS_RESX); goto dispatch_rmpp_rcv_done; default: /* Illegal return code */ CL_ASSERT(FALSE); } if (TRUE != is_last1) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "RMPP MADW assembly continues, TID=0x%" PRIx64 "\n", tid); goto dispatch_rmpp_rcv_done; } /* This is the last packet. */ if (0 == size) { /* The total size was not advertised in the first packet */ size = osmv_rmpp_recv_ctx_get_byte_num_from_last(p_recv_ctx); } /* NOTE: the received mad might not be >= 256 bytes. some MADs might contain several SA records but still be less then a full MAD. We have to use RMPP to send them over since on a regular "simple" MAD there is no way to know how many records were sent */ /* Build the MAD wrapper to be returned to the user. * The actual storage for the MAD is allocated there. */ p_new_madw = osm_mad_pool_get(p_bo->p_osm_pool, h_bind, size, p_mad_addr); if (NULL == p_new_madw) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_dispatch_rmpp_rcv: ERR 6506: " "Out Of Memory - could not allocate %d bytes for the MADW\n", size); status = IB_INSUFFICIENT_MEMORY; goto dispatch_rmpp_rcv_done; } p_req_madw = osmv_txn_get_madw(p_txn); p_mad_buf = osm_madw_get_mad_ptr(p_new_madw); status = osmv_rmpp_recv_ctx_reassemble_arbt_mad(p_recv_ctx, size, (uint8_t *) p_mad_buf); if (IB_SUCCESS != status) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_ERROR, "__osmv_dispatch_rmpp_rcv: ERR 6507: " "Internal error - could not reassemble the result MAD\n"); goto dispatch_rmpp_rcv_done; /* What can happen here? */ } /* The MAD is assembled, we are about to apply the callback. * Delete the transaction context, unless the transaction is double sided */ if (FALSE == osmv_txn_is_rmpp_init_by_peer(p_txn) || FALSE == osmv_mad_is_multi_resp(p_mad)) { osmv_txn_done(h_bind, key, FALSE); } osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "RMPP MADW %p assembly complete, TID=0x%" PRIx64 "\n", p_new_madw, tid); p_mad_buf->trans_id = cl_hton64(tid); osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "Restoring the original TID to 0x%" PRIx64 "\n", cl_ntoh64(p_mad_buf->trans_id)); /* Finally, do the job! */ p_bo->recv_cb(p_new_madw, p_bo->cb_context, p_req_madw); dispatch_rmpp_rcv_done: OSM_LOG_EXIT(p_bo->p_vendor->p_log); return status; }
static void __osmv_dispatch_rmpp_mad(IN osm_bind_handle_t h_bind, IN const ib_mad_t * p_mad, IN osmv_txn_ctx_t * p_txn, IN const osm_mad_addr_t * p_mad_addr) { ib_api_status_t status = IB_SUCCESS; uint64_t key = cl_ntoh64(p_mad->trans_id); boolean_t is_init_by_peer = FALSE; osmv_bind_obj_t *p_bo = (osmv_bind_obj_t *) h_bind; osm_madw_t *p_madw; OSM_LOG_ENTER(p_bo->p_vendor->p_log); if (NULL == p_txn) { if (FALSE == osmv_rmpp_is_data(p_mad) || FALSE == osmv_rmpp_is_first(p_mad)) { osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "The MAD does not match any transaction " "and does not start a sender-initiated RMPP transfer.\n"); goto dispatch_rmpp_mad_done; } /* IB Spec 13.6.2.2. This is a Sender Initiated Transfer. My peer is the requester and RMPP Sender. I am the RMPP Receiver. */ status = osmv_txn_init(h_bind, /*tid==key */ key, key, &p_txn); if (IB_SUCCESS != status) { goto dispatch_rmpp_mad_done; } is_init_by_peer = TRUE; osm_log(p_bo->p_vendor->p_log, OSM_LOG_DEBUG, "A new sender-initiated transfer (TID=0x%" PRIx64 ") started\n", key); } if (OSMV_TXN_RMPP_NONE == osmv_txn_get_rmpp_state(p_txn)) { /* Case 1: Fall through from above. * Case 2: When the transaction was initiated by me * (a single request MAD), there was an uncertainty * whether the reply will be RMPP. Now it's resolved, * since the reply is RMPP! */ status = osmv_txn_init_rmpp_receiver(h_bind, p_txn, is_init_by_peer); if (IB_SUCCESS != status) { goto dispatch_rmpp_mad_done; } } switch (osmv_txn_get_rmpp_state(p_txn)) { case OSMV_TXN_RMPP_RECEIVER: status = __osmv_dispatch_rmpp_rcv(h_bind, p_mad, p_txn, p_mad_addr); if (IB_SUCCESS != status) { if (FALSE == osmv_txn_is_rmpp_init_by_peer(p_txn)) { /* This is a requester, still waiting for the reply. Apply the callback */ /* update the status of the p_madw */ p_madw = osmv_txn_get_madw(p_txn); p_madw->status = status; p_bo->send_err_cb(p_bo->cb_context, p_madw); } /* ABORT/STOP/LOCAL ERROR */ osmv_txn_done(h_bind, osmv_txn_get_key(p_txn), FALSE); } break; case OSMV_TXN_RMPP_SENDER: __osmv_dispatch_rmpp_snd(h_bind, p_mad, p_txn, p_mad_addr); /* If an error happens here, it's the sender thread to cleanup the txn */ break; default: CL_ASSERT(FALSE); } dispatch_rmpp_mad_done: OSM_LOG_EXIT(p_bo->p_vendor->p_log); }