--- A/ulp/ipoib_NDIS6_CM/kernel/ipoib_port.h Tue Jan 11 17:54:42 2011 +++ B/ulp/ipoib_NDIS6_CM/kernel/ipoib_port.h Tue Jan 11 16:54:23 2011 @@ -72,6 +72,10 @@ #define IPOIB_RECV_FROM_NBL( P ) \ (((ipoib_recv_desc_t**)NET_BUFFER_LIST_MINIPORT_RESERVED(P))[0]) +//Used in CM RECV flow +#define IPOIB_CM_RECV_FROM_NBL( P ) \ + (((ipoib_cm_recv_desc_t**)NET_BUFFER_LIST_MINIPORT_RESERVED(P))[0]) + //Used in CM SEND flow - to update #define IPOIB_LIST_ITEM_FROM_NBL( P ) \ ((cl_list_item_t*)NET_BUFFER_LIST_MINIPORT_RESERVED(P)) @@ -102,6 +106,7 @@ ib_qp_handle_t h_qp; ib_query_handle_t h_query; ib_srq_handle_t h_srq; + atomic32_t srq_qp_cnt; net32_t qpn; ib_mr_handle_t h_mr; @@ -126,11 +131,20 @@ * Send CQ handle. * * h_qp -* QP handle for data transfers. +* QP handle for UD data transfers. * * h_query * Query handle for cancelling SA queries. * +* h_srq +* SRQ (Shared Receive Queue) handle +* +* srq_qp_cnt +* number of QPs bound to the SRQ +* +* qpn +* local QP number in net byte-order. +* * h_mr * Registration handle for all of physical memory. Used for * send/receive buffers to simplify growing the receive pool. @@ -138,6 +152,9 @@ * lkey * LKey for the memory region. * +* rate +* port rate +* * bcast_rec * Cached information about the broadcast group, used to specify * parameters used to join other multicast groups. @@ -337,8 +354,8 @@ * * send_buf_list * Lookaside list for dynamically allocating send buffers for send -* that require copies (ARP, DHCP, and any with more physical pages -* than can fit in the local data segments). +* operations which require copies (ARP, DHCP, IP fragmentation and any with more +* physical pages than can fit in the local data segments). *********/ typedef enum _ipoib_pkt_type @@ -350,31 +367,101 @@ } ipoib_pkt_type_t; -typedef struct _ipoib_cm_desc + +typedef enum __ib_recv_mode { - cl_pool_item_t item; /* Must be first. */ - uint32_t len; - ipoib_pkt_type_t type; - ib_recv_wr_t wr; - ib_local_ds_t local_ds[2]; - cl_list_item_t list_item; - uint8_t* p_alloc_buf; - uint8_t* p_buf; - uint32_t alloc_buf_size; - uint32_t buf_size; - net32_t lkey; - ib_mr_handle_t h_mr; - NDIS_TCP_IP_CHECKSUM_PACKET_INFO ndis_csum; + RECV_UD = 1, + RECV_RC = 2 + +} ib_recv_mode_t; -} ipoib_cm_desc_t; + + +typedef struct _ipoib_cm_recv_desc +{ + cl_pool_item_t item; /* Must be first. */ + uint32_t len; + ipoib_pkt_type_t type; + ib_recv_mode_t recv_mode;/* matches ipoib_recv_desc_t to this offset */ + ib_recv_wr_t wr; + ib_local_ds_t local_ds[MAX_CM_RECV_SGE]; /* 1 ds_local / phys page */ + cl_list_item_t list_item; + ipoib_endpt_t* p_endpt; + uint8_t* p_alloc_buf; + uint8_t* p_buf; + uint32_t alloc_buf_size; + uint32_t buf_size; + NET_BUFFER_LIST* p_NBL; + MDL* p_mdl; + NDIS_TCP_IP_CHECKSUM_NET_BUFFER_LIST_INFO csum; + +} ipoib_cm_recv_desc_t; +/* +* FIELDS +* item +* Pool item for storing descriptors in a pool. +* +* len +* Length to indicate to NDIS. This is different than the length of the +* received data as some data is IPoIB specific and filtered out. +* +* type +* Type of packet, used in filtering received packets against the packet +* filter. Also used to update stats. +* +* recv_mode +* IB connection mode RC(reliable connection) or UD(datagram) +* +* wr +* Receive work request. +* +* local_ds +* Local data segments. The second segment is only used if a buffer +* spans physical pages. +* +* list_item +* used when this recv desc is on the recv list (valid data) +* +* p_endpt +* remote endpoint struct +* +* p_alloc_buf +* allocated receive buffer start address; also start of Ethernet header. +* +* p_buf +* Buffer for the CM receive as p_aloc_buf+OFFSET == p_buf, due to adding Enet +* header in front of received ipoib CM data. ipoib_hdeader_t is overlayed +* by the synthesized ethernet header prior to passing buffer to NDIS receive +* routine. +* +* p_alloc_buf_size +* actual byte size of allocated buffer. +* +* buf_size +* byte size from p_buf to end of allocated buffer. +* +* p_NBL +* allocated NBL (NetworkBufferList) +* +* p_MDL +* allocated MDL (Memory Descriptor List) for p_buf - sizeof(eth_hdr) +* +* csum +* hardware utilized checksum status block. +*/ typedef struct _ipoib_recv_desc { cl_pool_item_t item; /* Must be first. */ uint32_t len; ipoib_pkt_type_t type; + ib_recv_mode_t recv_mode; /* matches ipoib_cm_recv_desc_t to this offset */ ib_recv_wr_t wr; ib_local_ds_t local_ds[2]; +#if UD_NBL_IN_DESC + NET_BUFFER_LIST *p_NBL; + MDL *p_mdl; +#endif NDIS_TCP_IP_CHECKSUM_PACKET_INFO ndis_csum; #if IPOIB_INLINE_RECV recv_buf_t buf; @@ -396,6 +483,9 @@ * Type of packet, used in filtering received packets against the packet * filter. Also used to update stats. * +* recv_mode +* IB recv mode: RC(reliable connection) or UD(datagram) +* * wr * Receive work request. * @@ -403,6 +493,9 @@ * Local data segments. The second segment is only used if a buffer * spans physical pages. * +* ndis_csum +* hardware utilized checksum status block. +* * buf * Buffer for the receive. * @@ -418,42 +511,28 @@ ib_local_ds_t local_ds[MAX_SEND_SGE]; /* Must be last. */ } ipoib_send_wr_t; -typedef enum __send_dir -{ - SEND_UD_QP = 1, - SEND_RC_QP = 2 -} send_dir_t; typedef struct _ipoib_send_desc { PNET_BUFFER_LIST p_netbuf_list; - ipoib_endpt_t *p_endpt; ib_qp_handle_t send_qp; - send_dir_t send_dir; uint32_t num_wrs; ipoib_send_wr_t send_wr[MAX_WRS_PER_MSG]; } ipoib_send_desc_t; /* * FIELDS -* p_pkt -* Pointer to the NDIS_PACKET associated with the send operation. -* -* p_endpt -* Endpoint for this send. +* p_netbuf_list +* Pointer to the NET_BUFFER_LIST associated with the send operation. * -* p_buf -* Buffer for the send, if allocated. -* -* wr -* Send work request. +* send_qp +* QP on which to Send work request. * -* pkt_hdr -* IPoIB packet header, pointed to by the first local datasegment. +* num_wrs +* count of work-requests (in send_wr) linked together. * -* local_ds -* Local data segment array. Placed last to allow allocating beyond the -* end of the descriptor for additional datasegments. +* send_wr +* Vector of IB work requests * * NOTES * The pool item is always first to allow casting form a cl_pool_item_t or @@ -478,7 +557,7 @@ * * done_list * List of receive descriptors (ipoib_desc_t) polled from the RX CQ which -* are used to construct the recv_NBL_array, which is then used to indicated +* are used to construct the recv_NBL_array; array is then used to indicate * received packets to NDIS 6. *********/ @@ -616,8 +695,7 @@ ipoib_send_desc_t * p_desc; ipoib_endpt_mgr_t endpt_mgr; - endpt_buf_mgr_t cm_buf_mgr; - endpt_recv_mgr_t cm_recv_mgr; + cm_buf_mgr_t cm_buf_mgr; ipoib_endpt_t *p_local_endpt; ib_ca_attr_t *p_ca_attrs; @@ -635,6 +713,7 @@ ib_net16_t base_lid; LONG n_no_progress; PIO_WORKITEM pPoWorkItem; + PIO_WORKITEM pPoWorkItemCM; ipoib_hdr_t hdr[1]; /* Must be last! */ } ipoib_port_t; @@ -653,10 +732,10 @@ * State of the port object. Tracks QP state fairly closely. * * recv_lock -* Spinlock to protect receive operations. +* Spinlock to protect UD receive operations. * * send_lock -* Spinlock to protect send operations. +* Spinlock to protect UD/RC send operations from NDIS callback. * * p_adapter * Parent adapter. Used to get AL handle. @@ -667,17 +746,73 @@ * max_sq_sge_supported * The actual number of SGEs that will be used for UD QP * +* sa_event +* Event signalled on SA query completion. +* +* mcast_cnt +* count of mcast group members +* +* leave_mcast_event +* event signalled upon leaving mcast group +* * ib_mgr * IB resource manager. * -* recv_mgr -* Receive manager. +* ipoib_buf_mgr +* send/recv buffer pools * * send_mgr * Send manager. * +* recv_mgr +* Recv manager. +* +* p_desc +* send descriptor used by all send operations (serialization point). +* * endpt_mgr -* Endpoint manager. +* Endpoint manager +* +* cm_buf_mgr +* Connected Mode SRQ recv buffer mgr. +* +* p_local_endpoint +* our local endpoint description +* p_ca_attrs +* HCA attributes for this port +* +* ref +* object reference count tracking +* endpt_rdr +* endpoint reader semaphore +* +* hdr_idx +* ipoib header array hdr[] index. +* +* pkey_index +* +* gc_dpc +* garbage collector DPC thread. +* +* gc_timer +* timer which starts periodic running of gc thread. +* +* bc_join_retry +* broadcast join_retry_cnt +* +* base_lid +* port's base LID +* +* n_no_progress +* +* pPoWorkItem +* DPC recv offload to worker thread +* +* pPoWorkItemCM +* DPC recv offload to worker thread for Connected Mode +* +* hdr +* ipoib header array - 1 entry per outstanding send: UD or RC; see hdr_idx. *********/ typedef struct _sgl_context { @@ -702,11 +837,44 @@ PNET_BUFFER_LIST p_nbl; PNET_BUFFER p_curr_nb; ipoib_endpt_t *p_endpt; + ipoib_send_desc_t *p_send_desc; send_buf_t *p_send_buf; PSCATTER_GATHER_LIST p_sgl; UINT tcp_payload; PVOID p_sg_buf; } ipoib_send_NB_SG; +/* +* FIELDS +* +* pool_item +* for storing descriptors in the send_mgr.send_pool. +* +* p_port +* HCA port which send is going out on. +* +* p_nbl +* Net buffer list - data passed down from NDIS to xmit +* +* p_cur_nb +* current list item in p_nbl list. +* +* p_endpt +* sending (aka local) endpoint & link to sending HCA port. +* +* *p_desc +* points to port->p_desc +* p_desc->p_endpt == destination EP. +* +* p_sgl +* pointer to a scatter-gather list +* +* tcp_payload +* used in LSO processing +* +* p_sg_buf +* scatter-gather buffer, used by p_sgl. +* +*********/ ib_api_status_t ipoib_create_port( @@ -730,14 +898,12 @@ ib_api_status_t ipoib_port_join_mcast( IN ipoib_port_t* const p_port, - IN const mac_addr_t mac, - IN const uint8_t state ); - + IN const mac_addr_t mac, + IN const uint8_t state ); void ipoib_leave_mcast_cb( - IN void *context ); - + IN void *context ); void ipoib_port_remove_endpt( @@ -823,19 +989,23 @@ ipoib_port_srq_destroy( IN ipoib_port_t* const p_port ); -#if 0 //CM +#if IPOIB_CM + ib_api_status_t ipoib_port_listen( IN ipoib_port_t* const p_port ); ib_api_status_t ipoib_port_cancel_listen( - IN ipoib_port_t* const p_port, - IN ipoib_endpt_t* const p_endpt ) { - UNUSED_PARAM(p_port); - UNUSED_PARAM(p_endpt); - return IB_SUCCESS; -} + IN ipoib_endpt_t* const p_endpt ); + +void +ipoib_send_complete_net_buffer( + IN ipoib_send_NB_SG *s_buf, + IN NDIS_STATUS status, + IN ULONG compl_flags, + IN boolean_t bLock ); + #endif ib_api_status_t @@ -846,20 +1016,6 @@ endpt_cm_buf_mgr_destroy( IN ipoib_port_t* const p_port ); -void -endpt_cm_buf_mgr_reset( - IN ipoib_port_t* const p_port ); - -void -endpt_cm_buf_mgr_put_recv( - IN endpt_buf_mgr_t * const p_buf_mgr, - IN ipoib_cm_desc_t* const p_desc ); - -void -endpt_cm_buf_mgr_put_recv_list( - IN endpt_buf_mgr_t * const p_buf_mgr, - IN cl_qlist_t* const p_list ); - uint32_t endpt_cm_recv_mgr_build_pkt_array( IN ipoib_port_t* const p_port, @@ -867,24 +1023,39 @@ IN cl_qlist_t* const p_done_list, IN OUT uint32_t* p_bytes_recv ); -ib_api_status_t -endpt_cm_post_recv( - IN ipoib_port_t* const p_port ); -/*void -endpt_cm_destroy_conn( +#if IPOIB_CM + +void +ipoib_cm_buf_mgr_put_recv( + IN ipoib_port_t* const p_port, + IN ipoib_cm_recv_desc_t* const p_desc, + IN NET_BUFFER_LIST* const p_NBL OPTIONAL ); + +ULONG ipoib_free_received_NBL( + IN ipoib_port_t *p_port, + IN NET_BUFFER_LIST *p_net_buffer_lists ); + +BOOLEAN +cm_destroy_conn( IN ipoib_port_t* const p_port, IN ipoib_endpt_t* const p_endpt ); -*/ + void endpt_cm_disconnect( - IN ipoib_port_t* const p_port, IN ipoib_endpt_t* const p_endpt ); + void -endpt_cm_flush_recv( +endpt_cm_release_resources( IN ipoib_port_t* const p_port, IN ipoib_endpt_t* const p_endpt ); +char * +get_ipoib_pkt_type_str( + IN ipoib_pkt_type_t t ); + +#endif // IPOIB_CM + ib_api_status_t ipoib_recv_dhcp( IN ipoib_port_t* const p_port, @@ -918,23 +1089,19 @@ } - - - // This function is only used to monitor send failures extern ULONG g_NBL_complete; extern ULONG g_NBL; static inline VOID NdisMSendNetBufferListsCompleteX( - IN ipoib_adapter_t* p_adapter, - IN PNET_BUFFER_LIST NetBufferLists, - IN ULONG SendCompleteFlags - ) + IN ipoib_adapter_t* p_adapter, + IN PNET_BUFFER_LIST NetBufferLists, + IN ULONG SendCompleteFlags ) { ++g_NBL_complete; ipoib_cnt_inc( &p_adapter->n_send_NBL_done ); -#ifdef DBG +#if DBG ASSERT(NET_BUFFER_LIST_NEXT_NBL(NetBufferLists) == NULL); if (NET_BUFFER_LIST_STATUS(NetBufferLists) != NDIS_STATUS_SUCCESS) { @@ -943,10 +1110,17 @@ NET_BUFFER_LIST_STATUS(NetBufferLists))); } IPOIB_PRINT( TRACE_LEVEL_VERBOSE, IPOIB_DBG_SEND, - ("Completing NBL=%x, g_NBL=%d, g_NBL_completed=%d \n", NetBufferLists, g_NBL, g_NBL_complete) ); + ("Completing NBL=%x, g_NBL=%d, g_NBL_completed=%d \n", + NetBufferLists, g_NBL, g_NBL_complete) ); #endif - NdisMSendNetBufferListsComplete(p_adapter->h_adapter,NetBufferLists,SendCompleteFlags); + NdisMSendNetBufferListsComplete( p_adapter->h_adapter, + NetBufferLists, + SendCompleteFlags ); } + +#if DBG +char *get_ib_recv_mode_str(ib_recv_mode_t m); +#endif #endif /* _IPOIB_PORT_H_ */