--- A/ulp/ipoib_NDIS6_CM/kernel/ipoib_endpoint.h Tue Jan 11 17:56:40 2011 +++ B/ulp/ipoib_NDIS6_CM/kernel/ipoib_endpoint.h Tue Jan 11 16:46:43 2011 @@ -46,37 +46,74 @@ #include "ipoib_debug.h" -typedef struct _endpt_buf_mgr +typedef struct _cm_buf_mgr { cl_qpool_t recv_pool; - NDIS_HANDLE h_packet_pool; - NDIS_HANDLE h_buffer_pool; - cl_qlist_t posted_list; + NDIS_HANDLE h_nbl_pool; + cl_spinlock_t lock; + cl_qlist_t oop_list; + long posted; + int32_t recv_pool_depth; boolean_t pool_init; -} endpt_buf_mgr_t; +} cm_buf_mgr_t; +/* +* FIELDS +* recv_pool +* recv descriptor pool - shared by all endpoints & posted to SRQ. +* +* h_nbl_pool +* handle to the pool of NDIS NETWORK_BUFFER_LISTs +* +* lock +* serialize access to the recv_pool & oop_list. +* +* oop_list +* list of recv pool buffers which are Out-Of-Pool. +* (SRQ bound QPs do not flush buffers on transition to QP Error state, this +* list tracks recv pool buffers by hand). +* +* posted +* Number of outstanding recv pool elements on the oop_list. +* Normally these pool elements are posted to the SRQ (for all endpoints) +* CM recv code wants to keep (posted == params.rq_depth). +* +* recv_pool_depth +* Total number of recv descriptor pool elements. +* +* pool_init +* boolean: TRUE == recv pool has been initialized. +*/ typedef struct _endpt_recv_mgr { - int32_t depth; - int32_t rq_depth; - //NDIS60 - //NDIS_PACKET **recv_NBL_array; - NET_BUFFER_LIST *recv_lst_array; + NET_BUFFER_LIST *NBL; + cl_qlist_t done_list; } endpt_recv_mgr_t; +/* +* FIELDS +* NBL +* Linked list of one or more NBL's (Network Buffer Lists) chained together. +* Each NBL points @ a ipoib_cm_desc_t->p_buf of RC received packet data. +* +* done_list +* list of completed WR (work Requests) which have passed recv filtering. +* Passed in that the WRs can be converted into NBLs and passed up to NDIS. +*/ typedef enum _cm_state { IPOIB_CM_DISCONNECTED, - IPOIB_CM_INIT, - IPOIB_CM_CONNECT, + IPOIB_CM_QUEUED_TO_CONNECT, + IPOIB_CM_CONNECTING, IPOIB_CM_CONNECTED, IPOIB_CM_LISTEN, IPOIB_CM_DREP_SENT, IPOIB_CM_DREQ_SENT, - IPOIB_CM_REJ_RECVD, + IPOIB_CM_DISCONNECT_CLEANUP, IPOIB_CM_DESTROY + } cm_state_t; typedef struct _cm_private_data @@ -90,15 +127,59 @@ ib_net64_t service_id; cm_private_data_t private_data; ib_qp_handle_t h_send_qp; - ib_qp_handle_t h_recv_qp; - ib_qp_handle_t h_work_qp; + ib_qp_handle_t h_send_qp_err; ib_cq_handle_t h_send_cq; + ib_qp_handle_t h_recv_qp; ib_cq_handle_t h_recv_cq; ib_listen_handle_t h_cm_listen; cm_state_t state; } endpt_conn_t; +/* +* FIELDS +* service_id +* listen() on this service ID +* +* private_data +* private data received from remote side. +* +* h_send_qp +* RC qp for send +* +* h_send_qp_err +* If !null, then copy of h_send_qp prior to setting h_send_qp == NULL. +* QP is in error state, awaiting destroy. +* +* h_recv_qp +* Rx RC qp handle +* +* h_send_cq +* Tx CQ handle +* +* h_recv_cq +* Rx CQ handle +* +* h_cm_listen +* listen()ing CM handle. +* +* state +* connection state for the active connection only. +* +* NOTES +* IpoIB Connect Mode (CM) connection protocol in a nutshell. +* An IPoIB interface encodes it CM capability in the hardware address it publishes. +* Once NDIS hands a Network Buffer to IPoIB for transmission (unicast), IPoIB needs +* to resolve the hardware address to IB adress information (LID, SL etc.). +* Once resolution is complete, an RC connection is forged with the remote host +* using IBAL CM. +* The connection process is symetrical - in that, for two communicating hosts +* there are two RC connections such that each connection works half duplex. +* CM state variable is that of the 'active' connection (h_send_qp), the passive +* side (h_recv_qp) setup does not alter the CM state variable. +* +*********/ + typedef struct _ipoib_endpt { cl_obj_t obj; @@ -111,30 +192,44 @@ ib_query_handle_t h_query; ib_mcast_handle_t h_mcast; mac_addr_t mac; - ib_gid_t dgid; net16_t dlid; + ib_gid_t dgid; net32_t qpn; - uint8_t cm_flag; ib_av_handle_t h_av; - endpt_conn_t conn; - ib_al_ifc_t *p_ifc; boolean_t is_in_use; boolean_t is_mcast_listener; + endpt_recv_mgr_t cm_recv; + endpt_conn_t conn; + uint32_t tx_mtu; + uint8_t cm_flag; + uint8_t cm_rx_flushing; + uint8_t cm_ep_destroy; + char tag[24]; // <(Broad/Multi)-cast or 0xLID> string + } ipoib_endpt_t; /* * FIELDS -* mac_item -* Map item for storing the endpoint in a map. The key is the -* destination MAC address. +* obj +* Pointer to the EndPoint object proper. * -* lid_item -* Map item for storing the endpoint in a map. The key is the -* destination LID. +* rel +* Object relations - used to convert from EndPoint to Port struct pointer. * * gid_item -* Map item for storing the endpoint in a map. The key is the -* destination GID. +* Map item for storing the endpoint in a map. key is destination GID. +* +* lid_item +* Map item for storing the endpoint in a map. key is destination LID. +* +* conn_item +* Map item for storing the endpoint in a connect map. key is hardware MAC. +* +* mac_item +* Map item for storing the endpoint in a map. key is destination MAC address. +* +* list_item +* used when emdpoint is on the connection list. * * h_query * Query handle for cancelling SA queries. @@ -143,31 +238,52 @@ * For multicast endpoints, the multicast handle. * * mac -* MAC address. -* -* dgid -* Destination GID. +* MAC address; next 2 bytes make for even alignment. * * dlid * Destination LID. The destination LID is only set for endpoints * that are on the same subnet. It is used as key in the LID map. * +* dgid +* Destination GID. +* * qpn -* Destination queue pair number. +* Destination UD queue pair number. * * h_av * Address vector for sending data. * -* expired -* Flag to indicate that the endpoint should be flushed. -* -* connection -* for connected mode endpoints -* * p_ifc * Reference to transport functions, can be used * while endpoint is not attached to port yet. * +* is_in_use +* Endpoint is a member of a mcast group. +* +* is_mcast_listener +* +* cm_recv +* Manage NDIS NBLs (Network Buffer List) and completed recv work-requests. +* +* conn +* for connected mode endpoints, IB RC connection info; includes conn state. +* +* tx_mtu +* current MTU; starts as UD MTU, when CM connected then +* tx_mtu = params.cm_payload_mtu, otherwise revert back to UD mtu. +* +* cm_flag +* != 0 implies CM capable. +* +* cm_rx_flushing +* != 0, CM is flushing SRQ QPs; defer object destroy until done flushing. +* +* cm_ep_destroy +* SRQ async error routine should destroy the EP object. +* +* tag +* Endpoint tag string: asciz string 'lid 0xNNN' +* * NOTES * If the h_mcast member is set, the endpoint is never expired. *********/ @@ -198,7 +314,7 @@ cl_obj_ref( &p_endpt->obj ); #if DBG IPOIB_PRINT( TRACE_LEVEL_INFORMATION, IPOIB_DBG_OBJ, - ("[%#x] Endpt refcount raised to %d\n", p_endpt, p_endpt->obj.ref_cnt)); + ("Endpt %s ++refcount %d\n", p_endpt->tag, p_endpt->obj.ref_cnt)); #endif /* * Anytime we reference the endpoint, we're either receiving data @@ -216,10 +332,12 @@ cl_obj_deref( &p_endpt->obj ); #if DBG IPOIB_PRINT( TRACE_LEVEL_INFORMATION, IPOIB_DBG_OBJ, - ("[%#x] Endpt refcount decremented to %d\n", p_endpt, p_endpt->obj.ref_cnt)); + ("Endpt %s --refcount %d\n", p_endpt->tag, p_endpt->obj.ref_cnt)); #endif } +void ipoib_endpt_cm_mgr_thread( + IN void* p_context ); NDIS_STATUS ipoib_endpt_queue( @@ -230,17 +348,21 @@ ipoib_endpt_parent( IN ipoib_endpt_t* const p_endpt ); -inline cm_state_t +void +endpt_unmap_conn_dgid( + IN ipoib_port_t* const p_port, + IN ipoib_endpt_t* const p_endpt ); + +static inline cm_state_t endpt_cm_set_state( IN ipoib_endpt_t* const p_endpt, IN cm_state_t state ) { - return (cm_state_t)InterlockedExchange( - (volatile LONG *)&p_endpt->conn.state, - (LONG)state ); + return (cm_state_t) InterlockedExchange( (volatile LONG *)&p_endpt->conn.state, + (LONG)state ); } -inline cm_state_t +static inline cm_state_t endpt_cm_get_state( IN ipoib_endpt_t* const p_endpt ) { @@ -250,19 +372,42 @@ } ib_api_status_t -endpt_cm_create_qp( - IN ipoib_endpt_t* const p_endpt, - IN ib_qp_handle_t* const p_h_qp ); +endpt_cm_connect( + IN ipoib_endpt_t* const p_endpt ); -ib_api_status_t -ipoib_endpt_connect( +void +endpt_queue_cm_connection( + IN ipoib_port_t* const p_port, IN ipoib_endpt_t* const p_endpt ); -int32_t -endpt_cm_recv_mgr_filter( +void +cm_release_resources( + IN ipoib_port_t* const p_port, IN ipoib_endpt_t* const p_endpt, - IN ib_wc_t* const p_done_wc_list, - OUT cl_qlist_t* const p_done_list, - OUT cl_qlist_t* const p_bad_list ); + IN int which_res ); + +void +cm_destroy_recv_resources( + IN ipoib_port_t* const p_port, + IN ipoib_endpt_t* const p_endpt ); + + +char * +cm_get_state_str( + IN cm_state_t ); + +char * +get_eth_packet_type_str( + IN net16_t); + +char * +get_IP_protocol_str( + IN uint8_t); + + +#if DBG + +void decode_enet_pkt( char *preFix, void *hdr, int len, char *postFix ); +#endif #endif /* _IPOIB_ENDPOINT_H_ */