X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fopeniblnd%2Fopeniblnd.h;h=c3e983701c9581f260c0823c7250dc34f441e2b6;hp=301d3ae051df84760d073b4c03a977b2533e9e42;hb=545fbb58e6fd30ffa5ca337e881a99625d8bf375;hpb=191061ee668400324f4505cf498f1ee2d57e4962 diff --git a/lnet/klnds/openiblnd/openiblnd.h b/lnet/klnds/openiblnd/openiblnd.h index 301d3ae..c3e9837 100644 --- a/lnet/klnds/openiblnd/openiblnd.h +++ b/lnet/klnds/openiblnd/openiblnd.h @@ -48,7 +48,10 @@ #include #include -#define DEBUG_SUBSYSTEM S_OPENIBNAL +#include +#include + +#define DEBUG_SUBSYSTEM S_NAL #include #include @@ -59,152 +62,180 @@ #include #include -#define OPENIBNAL_SERVICE_NAME "openibnal" - #if CONFIG_SMP -# define OPENIBNAL_N_SCHED num_online_cpus() /* # schedulers */ +# define IBNAL_N_SCHED num_online_cpus() /* # schedulers */ #else -# define OPENIBNAL_N_SCHED 1 /* # schedulers */ +# define IBNAL_N_SCHED 1 /* # schedulers */ #endif +#define IBNAL_N_CONND 4 /* # connection daemons */ -#define OPENIBNAL_MIN_RECONNECT_INTERVAL HZ /* first failed connection retry... */ -#define OPENIBNAL_MAX_RECONNECT_INTERVAL (60*HZ) /* ...exponentially increasing to this */ +#define IBNAL_MIN_RECONNECT_INTERVAL HZ /* first failed connection retry... */ +#define IBNAL_MAX_RECONNECT_INTERVAL (60*HZ) /* ...exponentially increasing to this */ -#define OPENIBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ +#define IBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ -#define OPENIBNAL_MSG_QUEUE_SIZE 8 /* # messages in-flight */ -#define OPENIBNAL_CREDIT_HIGHWATER 6 /* when to eagerly return credits */ -#define OPENIBNAL_RETRY 7 /* # times to retry */ -#define OPENIBNAL_RNR_RETRY 7 /* */ -#define OPENIBNAL_CM_RETRY 7 /* # times to retry connection */ -#define OPENIBNAL_FLOW_CONTROL 1 -#define OPENIBNAL_RESPONDER_RESOURCES 8 +#define IBNAL_MSG_QUEUE_SIZE 8 /* # messages/RDMAs in-flight */ +#define IBNAL_CREDIT_HIGHWATER 6 /* when to eagerly return credits */ +#define IBNAL_RETRY 7 /* # times to retry */ +#define IBNAL_RNR_RETRY 7 /* */ +#define IBNAL_CM_RETRY 7 /* # times to retry connection */ +#define IBNAL_FLOW_CONTROL 1 +#define IBNAL_RESPONDER_RESOURCES 8 -#define OPENIBNAL_NTX 64 /* # tx descs */ -#define OPENIBNAL_NTX_NBLK 256 /* # reserved tx descs */ +#define IBNAL_NTX 64 /* # tx descs */ +#define IBNAL_NTX_NBLK 256 /* # reserved tx descs */ -#define OPENIBNAL_PEER_HASH_SIZE 101 /* # peer lists */ +#define IBNAL_PEER_HASH_SIZE 101 /* # peer lists */ -#define OPENIBNAL_RESCHED 100 /* # scheduler loops before reschedule */ +#define IBNAL_RESCHED 100 /* # scheduler loops before reschedule */ -#define OPENIBNAL_CONCURRENT_PEERS 1000 /* # nodes all talking at once to me */ +#define IBNAL_CONCURRENT_PEERS 1000 /* # nodes all talking at once to me */ /* default vals for runtime tunables */ -#define OPENIBNAL_IO_TIMEOUT 50 /* default comms timeout (seconds) */ +#define IBNAL_IO_TIMEOUT 50 /* default comms timeout (seconds) */ +#define IBNAL_LISTENER_TIMEOUT 5 /* default listener timeout (seconds) */ +#define IBNAL_BACKLOG 127 /* default listener backlog */ +#define IBNAL_PORT 988 /* default listener port */ /************************/ /* derived constants... */ /* TX messages (shared by all connections) */ -#define OPENIBNAL_TX_MSGS (OPENIBNAL_NTX + OPENIBNAL_NTX_NBLK) -#define OPENIBNAL_TX_MSG_BYTES (OPENIBNAL_TX_MSGS * OPENIBNAL_MSG_SIZE) -#define OPENIBNAL_TX_MSG_PAGES ((OPENIBNAL_TX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) - -/* we may have up to 2 completions per transmit */ -#define OPENIBNAL_TX_CQ_ENTRIES (2*OPENIBNAL_TX_MSGS) +#define IBNAL_TX_MSGS (IBNAL_NTX + IBNAL_NTX_NBLK) +#define IBNAL_TX_MSG_BYTES (IBNAL_TX_MSGS * IBNAL_MSG_SIZE) +#define IBNAL_TX_MSG_PAGES ((IBNAL_TX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) /* RX messages (per connection) */ -#define OPENIBNAL_RX_MSGS OPENIBNAL_MSG_QUEUE_SIZE -#define OPENIBNAL_RX_MSG_BYTES (OPENIBNAL_RX_MSGS * OPENIBNAL_MSG_SIZE) -#define OPENIBNAL_RX_MSG_PAGES ((OPENIBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) +#define IBNAL_RX_MSGS IBNAL_MSG_QUEUE_SIZE +#define IBNAL_RX_MSG_BYTES (IBNAL_RX_MSGS * IBNAL_MSG_SIZE) +#define IBNAL_RX_MSG_PAGES ((IBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) -/* 1 completion per receive, per connection */ -#define OPENIBNAL_RX_CQ_ENTRIES (OPENIBNAL_RX_MSGS * OPENIBNAL_CONCURRENT_PEERS) +/* we may have up to 2 completions per transmit + + 1 completion per receive, per connection */ +#define IBNAL_CQ_ENTRIES ((2*IBNAL_TX_MSGS) + \ + (IBNAL_RX_MSGS * IBNAL_CONCURRENT_PEERS)) -#define OPENIBNAL_RDMA_BASE 0x0eeb0000 -#define OPENIBNAL_FMR 1 -#define OPENIBNAL_CKSUM 0 -//#define OPENIBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_PROCESS -#define OPENIBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_INTERRUPT +#define IBNAL_RDMA_BASE 0x0eeb0000 +#define IBNAL_FMR 1 +#define IBNAL_CKSUM 1 +//#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_PROCESS +#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_INTERRUPT typedef struct { - int koib_io_timeout; /* comms timeout (seconds) */ - struct ctl_table_header *koib_sysctl; /* sysctl interface */ -} koib_tunables_t; + int kib_io_timeout; /* comms timeout (seconds) */ + int kib_listener_timeout; /* listener's timeout */ + int kib_backlog; /* listenter's accept backlog */ + int kib_port; /* where the listener listens */ + struct ctl_table_header *kib_sysctl; /* sysctl interface */ +} kib_tunables_t; typedef struct { - int oibp_npages; /* # pages */ - int oibp_mapped; /* mapped? */ - __u64 oibp_vaddr; /* mapped region vaddr */ - __u32 oibp_lkey; /* mapped region lkey */ - __u32 oibp_rkey; /* mapped region rkey */ - struct ib_mr *oibp_handle; /* mapped region handle */ - struct page *oibp_pages[0]; -} koib_pages_t; + int ibp_npages; /* # pages */ + int ibp_mapped; /* mapped? */ + __u64 ibp_vaddr; /* mapped region vaddr */ + __u32 ibp_lkey; /* mapped region lkey */ + __u32 ibp_rkey; /* mapped region rkey */ + struct ib_mr *ibp_handle; /* mapped region handle */ + struct page *ibp_pages[0]; +} kib_pages_t; typedef struct { - int koib_init; /* initialisation state */ - __u64 koib_incarnation; /* which one am I */ - int koib_shutdown; /* shut down? */ - atomic_t koib_nthreads; /* # live threads */ - - __u64 koib_cm_service_id; /* service number I listen on */ - ptl_nid_t koib_nid; /* my NID */ - struct semaphore koib_nid_mutex; /* serialise NID ops */ - struct semaphore koib_nid_signal; /* signal completion */ - - rwlock_t koib_global_lock; /* stabilize peer/conn ops */ - - struct list_head *koib_peers; /* hash table of all my known peers */ - int koib_peer_hash_size; /* size of koib_peers */ - atomic_t koib_npeers; /* # peers extant */ - atomic_t koib_nconns; /* # connections extant */ - - struct list_head koib_connd_conns; /* connections to progress */ - struct list_head koib_connd_peers; /* peers waiting for a connection */ - wait_queue_head_t koib_connd_waitq; /* connection daemons sleep here */ - unsigned long koib_connd_waketime; /* when connd will wake */ - spinlock_t koib_connd_lock; /* serialise */ - - wait_queue_head_t koib_sched_waitq; /* schedulers sleep here */ - struct list_head koib_sched_txq; /* tx requiring attention */ - struct list_head koib_sched_rxq; /* rx requiring attention */ - spinlock_t koib_sched_lock; /* serialise */ + int kib_init; /* initialisation state */ + __u64 kib_incarnation; /* which one am I */ + int kib_shutdown; /* shut down? */ + atomic_t kib_nthreads; /* # live threads */ + + __u64 kib_svc_id; /* service number I listen on */ + tTS_IB_GID kib_svc_gid; /* device/port GID */ + __u16 kib_svc_pkey; /* device/port pkey */ + + ptl_nid_t kib_nid; /* my NID */ + struct semaphore kib_nid_mutex; /* serialise NID ops */ + struct semaphore kib_listener_signal; /* signal IP listener completion */ + struct socket *kib_listener_sock; /* IP listener's socket */ + int kib_listener_shutdown; /* ask IP listener to close */ + void *kib_listen_handle; /* IB listen handle */ + + rwlock_t kib_global_lock; /* stabilize peer/conn ops */ + + struct list_head *kib_peers; /* hash table of all my known peers */ + int kib_peer_hash_size; /* size of kib_peers */ + atomic_t kib_npeers; /* # peers extant */ + atomic_t kib_nconns; /* # connections extant */ + + struct list_head kib_reaper_conns; /* connections to reap */ + wait_queue_head_t kib_reaper_waitq; /* reaper sleeps here */ + unsigned long kib_reaper_waketime; /* when reaper will wake */ + spinlock_t kib_reaper_lock; /* serialise */ + + struct list_head kib_connd_peers; /* peers waiting for a connection */ + struct list_head kib_connd_acceptq; /* accepted sockets to handle */ + wait_queue_head_t kib_connd_waitq; /* connection daemons sleep here */ + spinlock_t kib_connd_lock; /* serialise */ + + wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */ + struct list_head kib_sched_txq; /* tx requiring attention */ + struct list_head kib_sched_rxq; /* rx requiring attention */ + spinlock_t kib_sched_lock; /* serialise */ - struct koib_tx *koib_tx_descs; /* all the tx descriptors */ - koib_pages_t *koib_tx_pages; /* premapped tx msg pages */ - - struct list_head koib_idle_txs; /* idle tx descriptors */ - struct list_head koib_idle_nblk_txs; /* idle reserved tx descriptors */ - wait_queue_head_t koib_idle_tx_waitq; /* block here for tx descriptor */ - __u64 koib_next_tx_cookie; /* RDMA completion cookie */ - spinlock_t koib_tx_lock; /* serialise */ + struct kib_tx *kib_tx_descs; /* all the tx descriptors */ + kib_pages_t *kib_tx_pages; /* premapped tx msg pages */ + + struct list_head kib_idle_txs; /* idle tx descriptors */ + struct list_head kib_idle_nblk_txs; /* idle reserved tx descriptors */ + wait_queue_head_t kib_idle_tx_waitq; /* block here for tx descriptor */ + __u64 kib_next_tx_cookie; /* RDMA completion cookie */ + spinlock_t kib_tx_lock; /* serialise */ - struct ib_device *koib_device; /* "the" device */ - struct ib_device_properties koib_device_props; /* its properties */ - int koib_port; /* port on the device */ - struct ib_port_properties koib_port_props; /* its properties */ - struct ib_pd *koib_pd; /* protection domain */ -#if OPENIBNAL_FMR - struct ib_fmr_pool *koib_fmr_pool; /* fast memory region pool */ + struct ib_device *kib_device; /* "the" device */ + struct ib_device_properties kib_device_props; /* its properties */ + int kib_port; /* port on the device */ + struct ib_port_properties kib_port_props; /* its properties */ + struct ib_pd *kib_pd; /* protection domain */ +#if IBNAL_FMR + struct ib_fmr_pool *kib_fmr_pool; /* fast memory region pool */ #endif - struct ib_cq *koib_rx_cq; /* receive completion queue */ - struct ib_cq *koib_tx_cq; /* transmit completion queue */ - void *koib_listen_handle; /* where I listen for connections */ - struct ib_common_attrib_service koib_service; /* SM service */ + struct ib_cq *kib_cq; /* completion queue */ -} koib_data_t; - -#define OPENIBNAL_INIT_NOTHING 0 -#define OPENIBNAL_INIT_DATA 1 -#define OPENIBNAL_INIT_LIB 2 -#define OPENIBNAL_INIT_PD 3 -#define OPENIBNAL_INIT_FMR 4 -#define OPENIBNAL_INIT_TXD 5 -#define OPENIBNAL_INIT_RX_CQ 6 -#define OPENIBNAL_INIT_TX_CQ 7 -#define OPENIBNAL_INIT_ALL 8 +} kib_data_t; + +#define IBNAL_INIT_NOTHING 0 +#define IBNAL_INIT_DATA 1 +#define IBNAL_INIT_LIB 2 +#define IBNAL_INIT_PD 3 +#define IBNAL_INIT_FMR 4 +#define IBNAL_INIT_TXD 5 +#define IBNAL_INIT_CQ 6 +#define IBNAL_INIT_ALL 7 + +typedef struct kib_acceptsock /* accepted socket queued for connd */ +{ + struct list_head ibas_list; /* queue for attention */ + struct socket *ibas_sock; /* the accepted socket */ +} kib_acceptsock_t; /************************************************************************ - * Wire message structs. + * IB Wire message format. * These are sent in sender's byte order (i.e. receiver flips). - * CAVEAT EMPTOR: other structs communicated between nodes (e.g. MAD - * private data and SM service info), is LE on the wire. + * They may be sent via TCP/IP (service ID,GID,PKEY query/response), + * as private data in the connection request/response, or "normally". */ +typedef struct kib_svcrsp /* service response */ +{ + __u64 ibsr_svc_id; /* service's id */ + __u8 ibsr_svc_gid[16]; /* service's gid */ + __u16 ibsr_svc_pkey; /* service's pkey */ +} kib_svcrsp_t; + +typedef struct kib_connparams +{ + __u32 ibcp_queue_depth; +} kib_connparams_t; + typedef struct { union { @@ -214,125 +245,125 @@ typedef struct __u32 md_lkey; __u32 md_rkey; __u64 md_addr; -} koib_md_t; +} kib_md_t; typedef struct { - __u32 rd_key; /* remote key */ - __u32 rd_nob; /* # of bytes */ - __u64 rd_addr; /* remote io vaddr */ -} koib_rdma_desc_t; - + __u32 rd_key; /* remote key */ + __u32 rd_nob; /* # of bytes */ + __u64 rd_addr; /* remote io vaddr */ +} kib_rdma_desc_t; typedef struct { - ptl_hdr_t oibim_hdr; /* portals header */ - char oibim_payload[0]; /* piggy-backed payload */ -} koib_immediate_msg_t; + ptl_hdr_t ibim_hdr; /* portals header */ + char ibim_payload[0]; /* piggy-backed payload */ +} kib_immediate_msg_t; typedef struct { - ptl_hdr_t oibrm_hdr; /* portals header */ - __u64 oibrm_cookie; /* opaque completion cookie */ - koib_rdma_desc_t oibrm_desc; /* where to suck/blow */ -} koib_rdma_msg_t; + ptl_hdr_t ibrm_hdr; /* portals header */ + __u64 ibrm_cookie; /* opaque completion cookie */ + kib_rdma_desc_t ibrm_desc; /* where to suck/blow */ +} kib_rdma_msg_t; typedef struct { - __u64 oibcm_cookie; /* opaque completion cookie */ - __u32 oibcm_status; /* completion status */ -} koib_completion_msg_t; + __u64 ibcm_cookie; /* opaque completion cookie */ + __u32 ibcm_status; /* completion status */ +} kib_completion_msg_t; typedef struct { - __u32 oibm_magic; /* I'm an openibnal message */ - __u16 oibm_version; /* this is my version number */ - __u8 oibm_type; /* msg type */ - __u8 oibm_credits; /* returned credits */ -#if OPENIBNAL_CKSUM - __u32 oibm_nob; - __u32 oibm_cksum; -#endif + /* First 2 fields fixed FOR ALL TIME */ + __u32 ibm_magic; /* I'm an openibnal message */ + __u16 ibm_version; /* this is my version number */ + + __u8 ibm_type; /* msg type */ + __u8 ibm_credits; /* returned credits */ + __u32 ibm_nob; /* # bytes in whole message */ + __u32 ibm_cksum; /* checksum (0 == no checksum) */ + __u64 ibm_srcnid; /* sender's NID */ + __u64 ibm_srcstamp; /* sender's incarnation */ + __u64 ibm_dstnid; /* destination's NID */ + __u64 ibm_dststamp; /* destination's incarnation */ union { - koib_immediate_msg_t immediate; - koib_rdma_msg_t rdma; - koib_completion_msg_t completion; - } oibm_u; -} koib_msg_t; - -#define OPENIBNAL_MSG_MAGIC 0x0be91b91 /* unique magic */ -#define OPENIBNAL_MSG_VERSION 1 /* current protocol version */ - -#define OPENIBNAL_MSG_NOOP 0xd0 /* nothing (just credits) */ -#define OPENIBNAL_MSG_IMMEDIATE 0xd1 /* portals hdr + payload */ -#define OPENIBNAL_MSG_PUT_RDMA 0xd2 /* portals PUT hdr + source rdma desc */ -#define OPENIBNAL_MSG_PUT_DONE 0xd3 /* signal PUT rdma completion */ -#define OPENIBNAL_MSG_GET_RDMA 0xd4 /* portals GET hdr + sink rdma desc */ -#define OPENIBNAL_MSG_GET_DONE 0xd5 /* signal GET rdma completion */ + kib_svcrsp_t svcrsp; + kib_connparams_t connparams; + kib_immediate_msg_t immediate; + kib_rdma_msg_t rdma; + kib_completion_msg_t completion; + } ibm_u; +} kib_msg_t; + +#define IBNAL_MSG_MAGIC 0x0be91b91 /* unique magic */ +#define IBNAL_MSG_VERSION 2 /* current protocol version */ + +#define IBNAL_MSG_SVCQRY 0xb0 /* service query */ +#define IBNAL_MSG_SVCRSP 0xb1 /* service response */ +#define IBNAL_MSG_CONNREQ 0xc0 /* connection request */ +#define IBNAL_MSG_CONNACK 0xc1 /* connection acknowledge */ +#define IBNAL_MSG_NOOP 0xd0 /* nothing (just credits) */ +#define IBNAL_MSG_IMMEDIATE 0xd1 /* portals hdr + payload */ +#define IBNAL_MSG_PUT_RDMA 0xd2 /* portals PUT hdr + source rdma desc */ +#define IBNAL_MSG_PUT_DONE 0xd3 /* signal PUT rdma completion */ +#define IBNAL_MSG_GET_RDMA 0xd4 /* portals GET hdr + sink rdma desc */ +#define IBNAL_MSG_GET_DONE 0xd5 /* signal GET rdma completion */ /***********************************************************************/ -typedef struct koib_rx /* receive message */ +typedef struct kib_rx /* receive message */ { struct list_head rx_list; /* queue for attention */ - struct koib_conn *rx_conn; /* owning conn */ + struct kib_conn *rx_conn; /* owning conn */ int rx_rdma; /* RDMA completion posted? */ int rx_posted; /* posted? */ __u64 rx_vaddr; /* pre-mapped buffer (hca vaddr) */ - koib_msg_t *rx_msg; /* pre-mapped buffer (host vaddr) */ + kib_msg_t *rx_msg; /* pre-mapped buffer (host vaddr) */ struct ib_receive_param rx_sp; /* receive work item */ struct ib_gather_scatter rx_gl; /* and it's memory */ -} koib_rx_t; +} kib_rx_t; -typedef struct koib_tx /* transmit message */ +typedef struct kib_tx /* transmit message */ { struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */ int tx_isnblk; /* I'm reserved for non-blocking sends */ - struct koib_conn *tx_conn; /* owning conn */ + struct kib_conn *tx_conn; /* owning conn */ int tx_mapped; /* mapped for RDMA? */ int tx_sending; /* # tx callbacks outstanding */ int tx_status; /* completion status */ - int tx_passive_rdma; /* waiting for peer to RDMA? */ - int tx_passive_rdma_wait; /* on ibc_rdma_queue */ - unsigned long tx_passive_rdma_deadline; /* completion deadline */ + unsigned long tx_deadline; /* completion deadline */ + int tx_passive_rdma; /* peer sucks/blows */ + int tx_passive_rdma_wait; /* waiting for peer to complete */ __u64 tx_passive_rdma_cookie; /* completion cookie */ lib_msg_t *tx_libmsg[2]; /* lib msgs to finalize on completion */ - koib_md_t tx_md; /* RDMA mapping (active/passive) */ + kib_md_t tx_md; /* RDMA mapping (active/passive) */ __u64 tx_vaddr; /* pre-mapped buffer (hca vaddr) */ - koib_msg_t *tx_msg; /* pre-mapped buffer (host vaddr) */ + kib_msg_t *tx_msg; /* pre-mapped buffer (host vaddr) */ int tx_nsp; /* # send work items */ struct ib_send_param tx_sp[2]; /* send work items... */ struct ib_gather_scatter tx_gl[2]; /* ...and their memory */ -} koib_tx_t; +} kib_tx_t; -#define KOIB_TX_UNMAPPED 0 -#define KOIB_TX_MAPPED 1 -#define KOIB_TX_MAPPED_FMR 2 +#define KIB_TX_UNMAPPED 0 +#define KIB_TX_MAPPED 1 +#define KIB_TX_MAPPED_FMR 2 -typedef struct koib_wire_connreq -{ - __u32 wcr_magic; /* I'm an openibnal connreq */ - __u16 wcr_version; /* this is my version number */ - __u16 wcr_queue_depth; /* this is my receive queue size */ - __u64 wcr_nid; /* peer's NID */ - __u64 wcr_incarnation; /* peer's incarnation */ -} koib_wire_connreq_t; - -typedef struct koib_connreq +typedef struct kib_connreq { - /* connection-in-progress */ - struct koib_conn *cr_conn; - koib_wire_connreq_t cr_wcr; - __u64 cr_tid; - struct ib_common_attrib_service cr_service; - tTS_IB_GID cr_gid; - struct ib_path_record cr_path; - struct ib_cm_active_param cr_connparam; -} koib_connreq_t; - -typedef struct koib_conn + /* active connection-in-progress state */ + struct kib_conn *cr_conn; + kib_msg_t cr_msg; + __u64 cr_tid; + tTS_IB_GID cr_gid; + kib_svcrsp_t cr_svcrsp; + struct ib_path_record cr_path; + struct ib_cm_active_param cr_connparam; +} kib_connreq_t; + +typedef struct kib_conn { - struct koib_peer *ibc_peer; /* owning peer */ + struct kib_peer *ibc_peer; /* owning peer */ struct list_head ibc_list; /* stash on peer's conn list */ __u64 ibc_incarnation; /* which instance of the peer */ atomic_t ibc_refcount; /* # users */ @@ -342,28 +373,31 @@ typedef struct koib_conn int ibc_credits; /* # credits I have */ int ibc_outstanding_credits; /* # credits to return */ struct list_head ibc_tx_queue; /* send queue */ - struct list_head ibc_rdma_queue; /* tx awaiting RDMA completion */ + struct list_head ibc_active_txs; /* active tx awaiting completion */ spinlock_t ibc_lock; /* serialise */ - koib_rx_t *ibc_rxs; /* the rx descs */ - koib_pages_t *ibc_rx_pages; /* premapped rx msg pages */ + kib_rx_t *ibc_rxs; /* the rx descs */ + kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */ struct ib_qp *ibc_qp; /* queue pair */ __u32 ibc_qpn; /* queue pair number */ tTS_IB_CM_COMM_ID ibc_comm_id; /* connection ID? */ - koib_connreq_t *ibc_connreq; /* connection request state */ -} koib_conn_t; + kib_connreq_t *ibc_connreq; /* connection request state */ +} kib_conn_t; -#define OPENIBNAL_CONN_INIT_NOTHING 0 /* initial state */ -#define OPENIBNAL_CONN_INIT_QP 1 /* ibc_qp set up */ -#define OPENIBNAL_CONN_CONNECTING 2 /* started to connect */ -#define OPENIBNAL_CONN_ESTABLISHED 3 /* connection established */ -#define OPENIBNAL_CONN_DEATHROW 4 /* waiting to be closed */ -#define OPENIBNAL_CONN_ZOMBIE 5 /* waiting to be freed */ +#define IBNAL_CONN_INIT_NOTHING 0 /* initial state */ +#define IBNAL_CONN_INIT_QP 1 /* ibc_qp set up */ +#define IBNAL_CONN_CONNECTING 2 /* started to connect */ +#define IBNAL_CONN_ESTABLISHED 3 /* connection established */ +#define IBNAL_CONN_DEATHROW 4 /* waiting to be closed */ +#define IBNAL_CONN_ZOMBIE 5 /* waiting to be freed */ -typedef struct koib_peer +typedef struct kib_peer { struct list_head ibp_list; /* stash on global peer list */ - struct list_head ibp_connd_list; /* schedule on koib_connd_peers */ + struct list_head ibp_connd_list; /* schedule on kib_connd_peers */ ptl_nid_t ibp_nid; /* who's on the other end(s) */ + __u32 ibp_ip; /* IP to query for peer conn params */ + int ibp_port; /* port to qery for peer conn params */ + __u64 ibp_incarnation; /* peer's incarnation */ atomic_t ibp_refcount; /* # users */ int ibp_persistence; /* "known" peer refs */ struct list_head ibp_conns; /* all active connections */ @@ -371,30 +405,29 @@ typedef struct koib_peer int ibp_connecting; /* connecting+accepting */ unsigned long ibp_reconnect_time; /* when reconnect may be attempted */ unsigned long ibp_reconnect_interval; /* exponential backoff */ -} koib_peer_t; +} kib_peer_t; - -extern lib_nal_t koibnal_lib; -extern koib_data_t koibnal_data; -extern koib_tunables_t koibnal_tunables; +extern lib_nal_t kibnal_lib; +extern kib_data_t kibnal_data; +extern kib_tunables_t kibnal_tunables; static inline struct list_head * -koibnal_nid2peerlist (ptl_nid_t nid) +kibnal_nid2peerlist (ptl_nid_t nid) { - unsigned int hash = ((unsigned int)nid) % koibnal_data.koib_peer_hash_size; + unsigned int hash = ((unsigned int)nid) % kibnal_data.kib_peer_hash_size; - return (&koibnal_data.koib_peers [hash]); + return (&kibnal_data.kib_peers [hash]); } static inline int -koibnal_peer_active(koib_peer_t *peer) +kibnal_peer_active(kib_peer_t *peer) { /* Am I in the peer hash table? */ return (!list_empty(&peer->ibp_list)); } static inline void -koibnal_queue_tx_locked (koib_tx_t *tx, koib_conn_t *conn) +kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) { /* CAVEAT EMPTOR: tx takes caller's ref on conn */ @@ -402,40 +435,13 @@ koibnal_queue_tx_locked (koib_tx_t *tx, koib_conn_t *conn) LASSERT (tx->tx_conn == NULL); /* only set here */ tx->tx_conn = conn; + tx->tx_deadline = jiffies + kibnal_tunables.kib_io_timeout * HZ; list_add_tail(&tx->tx_list, &conn->ibc_tx_queue); } -#define KOIBNAL_SERVICE_KEY_MASK (IB_SA_SERVICE_COMP_MASK_NAME | \ - IB_SA_SERVICE_COMP_MASK_DATA8_1 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_2 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_3 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_4 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_5 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_6 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_7 | \ - IB_SA_SERVICE_COMP_MASK_DATA8_8) - -static inline __u64* -koibnal_service_nid_field(struct ib_common_attrib_service *srv) -{ - /* must be consistent with KOIBNAL_SERVICE_KEY_MASK */ - return (__u64 *)srv->service_data8; -} - - -static inline void -koibnal_set_service_keys(struct ib_common_attrib_service *srv, ptl_nid_t nid) -{ - LASSERT (strlen (OPENIBNAL_SERVICE_NAME) < sizeof(srv->service_name)); - memset (srv->service_name, 0, sizeof(srv->service_name)); - strcpy (srv->service_name, OPENIBNAL_SERVICE_NAME); - - *koibnal_service_nid_field(srv) = cpu_to_le64(nid); -} - #if 0 static inline void -koibnal_show_rdma_attr (koib_conn_t *conn) +kibnal_show_rdma_attr (kib_conn_t *conn) { struct ib_qp_attribute qp_attr; int rc; @@ -455,54 +461,99 @@ koibnal_show_rdma_attr (koib_conn_t *conn) } #endif -#if CONFIG_X86 static inline __u64 -koibnal_page2phys (struct page *p) +kibnal_page2phys (struct page *p) { - __u64 page_number = p - mem_map; - - return (page_number << PAGE_SHIFT); + return page_to_phys(p); } -#else -# error "no page->phys" + +/* CAVEAT EMPTOR: + * We rely on tx/rx descriptor alignment to allow us to use the lowest bit + * of the work request id as a flag to determine if the completion is for a + * transmit or a receive. It seems that that the CQ entry's 'op' field + * isn't always set correctly on completions that occur after QP teardown. */ + +static inline __u64 +kibnal_ptr2wreqid (void *ptr, int isrx) +{ + unsigned long lptr = (unsigned long)ptr; + + LASSERT ((lptr & 1) == 0); + return (__u64)(lptr | (isrx ? 1 : 0)); +} + +static inline void * +kibnal_wreqid2ptr (__u64 wreqid) +{ + return (void *)(((unsigned long)wreqid) & ~1UL); +} + +static inline int +kibnal_wreqid_is_rx (__u64 wreqid) +{ + return (wreqid & 1) != 0; +} + +#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)) +# define sk_allocation allocation +# define sk_data_ready data_ready +# define sk_write_space write_space +# define sk_user_data user_data +# define sk_prot prot +# define sk_sndbuf sndbuf +# define sk_socket socket +# define sk_wmem_queued wmem_queued +# define sk_err err +# define sk_sleep sleep #endif -extern koib_peer_t *koibnal_create_peer (ptl_nid_t nid); -extern void koibnal_put_peer (koib_peer_t *peer); -extern int koibnal_del_peer (ptl_nid_t nid, int single_share); -extern koib_peer_t *koibnal_find_peer_locked (ptl_nid_t nid); -extern void koibnal_unlink_peer_locked (koib_peer_t *peer); -extern int koibnal_close_stale_conns_locked (koib_peer_t *peer, +extern void kibnal_init_msg(kib_msg_t *msg, int type, int body_nob); +extern void kibnal_pack_msg(kib_msg_t *msg, int credits, + ptl_nid_t dstnid, __u64 dststamp); +extern int kibnal_unpack_msg(kib_msg_t *msg, int nob); +extern void kibnal_handle_svcqry (struct socket *sock); +extern int kibnal_make_svcqry (kib_conn_t *conn); +extern void kibnal_free_acceptsock (kib_acceptsock_t *as); +extern int kibnal_listener_procint(ctl_table *table, int write, + struct file *filp, void *buffer, + size_t *lenp); +extern kib_peer_t *kibnal_create_peer (ptl_nid_t nid); +extern void kibnal_put_peer (kib_peer_t *peer); +extern int kibnal_del_peer (ptl_nid_t nid, int single_share); +extern kib_peer_t *kibnal_find_peer_locked (ptl_nid_t nid); +extern void kibnal_unlink_peer_locked (kib_peer_t *peer); +extern int kibnal_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation); -extern koib_conn_t *koibnal_create_conn (void); -extern void koibnal_put_conn (koib_conn_t *conn); -extern void koibnal_destroy_conn (koib_conn_t *conn); -extern int koibnal_alloc_pages (koib_pages_t **pp, int npages, int access); -extern void koibnal_free_pages (koib_pages_t *p); +extern kib_conn_t *kibnal_create_conn (void); +extern void kibnal_put_conn (kib_conn_t *conn); +extern void kibnal_destroy_conn (kib_conn_t *conn); +extern int kibnal_alloc_pages (kib_pages_t **pp, int npages, int access); +extern void kibnal_free_pages (kib_pages_t *p); -extern void koibnal_check_sends (koib_conn_t *conn); +extern void kibnal_check_sends (kib_conn_t *conn); extern tTS_IB_CM_CALLBACK_RETURN -koibnal_conn_callback (tTS_IB_CM_EVENT event, tTS_IB_CM_COMM_ID cid, +kibnal_conn_callback (tTS_IB_CM_EVENT event, tTS_IB_CM_COMM_ID cid, void *param, void *arg); extern tTS_IB_CM_CALLBACK_RETURN -koibnal_passive_conn_callback (tTS_IB_CM_EVENT event, tTS_IB_CM_COMM_ID cid, +kibnal_passive_conn_callback (tTS_IB_CM_EVENT event, tTS_IB_CM_COMM_ID cid, void *param, void *arg); -extern void koibnal_close_conn_locked (koib_conn_t *conn, int error); -extern void koibnal_destroy_conn (koib_conn_t *conn); -extern int koibnal_thread_start (int (*fn)(void *arg), void *arg); -extern int koibnal_scheduler(void *arg); -extern int koibnal_connd (void *arg); -extern void koibnal_rx_callback (struct ib_cq *cq, struct ib_cq_entry *e, void *arg); -extern void koibnal_tx_callback (struct ib_cq *cq, struct ib_cq_entry *e, void *arg); -extern void koibnal_init_tx_msg (koib_tx_t *tx, int type, int body_nob); -extern int koibnal_close_conn (koib_conn_t *conn, int why); -extern void koibnal_start_active_rdma (int type, int status, - koib_rx_t *rx, lib_msg_t *libmsg, - unsigned int niov, - struct iovec *iov, ptl_kiov_t *kiov, - size_t offset, size_t nob); +extern void kibnal_close_conn_locked (kib_conn_t *conn, int error); +extern void kibnal_destroy_conn (kib_conn_t *conn); +extern int kibnal_thread_start (int (*fn)(void *arg), void *arg); +extern int kibnal_scheduler(void *arg); +extern int kibnal_connd (void *arg); +extern int kibnal_reaper (void *arg); +extern void kibnal_callback (struct ib_cq *cq, struct ib_cq_entry *e, void *arg); +extern void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob); +extern int kibnal_close_conn (kib_conn_t *conn, int why); +extern void kibnal_start_active_rdma (int type, int status, + kib_rx_t *rx, lib_msg_t *libmsg, + unsigned int niov, + struct iovec *iov, ptl_kiov_t *kiov, + int offset, int nob); +