Whamcloud - gitweb
Landing b_hd_newconfig on HEAD
[fs/lustre-release.git] / lnet / klnds / openiblnd / openiblnd.h
index 5ba102e..6ed306c 100644 (file)
 #include <net/sock.h>
 #include <linux/in.h>
 
-#define DEBUG_SUBSYSTEM S_NAL
+#define DEBUG_SUBSYSTEM S_LND
 
 #include <libcfs/kp30.h>
-#include <portals/p30.h>
-#include <portals/lib-p30.h>
-#include <portals/nal.h>
+#include <lnet/lnet.h>
+#include <lnet/lib-lnet.h>
 
 #include <ts_ib_core.h>
 #include <ts_ib_cm.h>
 #include <ts_ib_sa_client.h>
 
+#ifndef USING_TSAPI
+
+/* OpenIB Gen1 */
+typedef struct ib_qp       ib_qp_t;
+typedef struct ib_mr       ib_mr_t;
+typedef struct ib_fmr      ib_fmr_t;
+typedef struct ib_pd       ib_pd_t;
+typedef struct ib_cq       ib_cq_t;
+typedef struct ib_fmr_pool ib_fmr_pool_t;
+
+#else
+
+/* Cisco (topspin) */
+typedef void                 ib_qp_t;
+typedef void                 ib_mr_t;
+typedef void                 ib_fmr_t;
+typedef void                 ib_pd_t;
+typedef void                 ib_cq_t;
+typedef void                 ib_fmr_pool_t;
+
+#define IB_ACCESS_LOCAL_WRITE              TS_IB_ACCESS_LOCAL_WRITE
+#define IB_WQ_SIGNAL_SELECTABLE            TS_IB_ACCESS_LOCAL_WRITE
+#define IB_TRANSPORT_RC                    TS_IB_TRANSPORT_RC
+#define IB_QP_STATE_INIT                   TS_IB_QP_STATE_INIT
+#define IB_QP_ATTRIBUTE_STATE              TS_IB_QP_ATTRIBUTE_STATE
+#define IB_QP_ATTRIBUTE_PORT               TS_IB_QP_ATTRIBUTE_PORT
+#define IB_QP_ATTRIBUTE_PKEY_INDEX         TS_IB_QP_ATTRIBUTE_PKEY_INDEX
+#define IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE
+#define IB_ACCESS_LOCAL_WRITE              TS_IB_ACCESS_LOCAL_WRITE
+#define IB_ACCESS_REMOTE_WRITE             TS_IB_ACCESS_REMOTE_WRITE
+#define IB_ACCESS_REMOTE_READ              TS_IB_ACCESS_REMOTE_READ
+#define IB_CQ_CALLBACK_INTERRU             TS_IB_CQ_CALLBACK_INTERRUPTPT
+#define IB_CQ_PROVIDER_REARM               TS_IB_CQ_PROVIDER_REARM
+#define IB_CQ_CALLBACK_INTERRUPT           TS_IB_CQ_CALLBACK_INTERRUPT
+#define IB_COMPLETION_STATUS_SUCCESS       TS_IB_COMPLETION_STATUS_SUCCESS
+#define IB_OP_SEND                         TS_IB_OP_SEND
+#define IB_OP_RDMA_WRITE                   TS_IB_OP_RDMA_WRITE
+#define IB_OP_RDMA_READ                    TS_IB_OP_RDMA_READ
+
+#endif
+
 #if CONFIG_SMP
 # define IBNAL_N_SCHED      num_online_cpus()   /* # schedulers */
 #else
 # define IBNAL_N_SCHED      1                   /* # schedulers */
 #endif
-#define IBNAL_N_CONND       4                   /* # connection daemons */
-
-#define IBNAL_MIN_RECONNECT_INTERVAL HZ         /* first failed connection retry... */
-#define IBNAL_MAX_RECONNECT_INTERVAL (60*HZ)    /* ...exponentially increasing to this */
-
-#define IBNAL_MSG_SIZE           (4<<10)        /* max size of queued messages (inc hdr) */
 
-#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 IBNAL_NTX                 64            /* # tx descs */
-#define IBNAL_NTX_NBLK            256           /* # reserved tx descs */
-
-#define IBNAL_PEER_HASH_SIZE      101           /* # peer lists */
+#define IBNAL_FMR                    1
+//#define IBNAL_CALLBACK_CTXT  IB_CQ_CALLBACK_PROCESS
+#define IBNAL_CALLBACK_CTXT  IB_CQ_CALLBACK_INTERRUPT
 
-#define IBNAL_RESCHED             100           /* # scheduler loops before reschedule */
 
-#define IBNAL_CONCURRENT_PEERS    1000          /* # nodes all talking at once to me */
+/* tunables fixed at compile time */
+#define IBNAL_PEER_HASH_SIZE         101        /* # peer lists */
+#define IBNAL_RESCHED                100        /* # scheduler loops before reschedule */
+#define IBNAL_MSG_QUEUE_SIZE         8          /* # messages/RDMAs in-flight */
+#define IBNAL_CREDIT_HIGHWATER       6          /* when to eagerly return credits */
+#define IBNAL_MSG_SIZE              (4<<10)     /* max size of queued messages (inc hdr) */
+#define IBNAL_RDMA_BASE              0x0eeb0000
 
-/* default vals for runtime tunables */
-#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 */
+/* QP tunables */
+#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
 
 /************************/
 /* derived constants... */
 
 /* TX messages (shared by all connections) */
-#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)
+#define IBNAL_TX_MSGS()       (*kibnal_tunables.kib_ntx)
+#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 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)
+#define IBNAL_RX_MSGS         (IBNAL_MSG_QUEUE_SIZE * 2)
+#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)
 
 /* 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 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
+#define IBNAL_CQ_ENTRIES()  ((2*IBNAL_TX_MSGS()) +                                      \
+                             (IBNAL_RX_MSGS * *kibnal_tunables.kib_concurrent_peers))
 
 typedef struct
 {
-        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 */
+        char    **kib_ipif_basename;            /* IPoIB interface base name */
+        int      *kib_n_connd;                  /* # connection daemons */
+        int      *kib_min_reconnect_interval;   /* min connect retry seconds... */
+        int      *kib_max_reconnect_interval;   /* max connect retry seconds */
+        int      *kib_concurrent_peers;         /* max # peers */
+        int      *kib_cksum;                    /* checksum kib_msg_t? */
+        int      *kib_timeout;                  /* comms timeout (seconds) */
+        int      *kib_keepalive;                /* keepalive (seconds) */
+        int      *kib_ntx;                      /* # tx descs */
+        int      *kib_credits;                  /* # concurrent sends */
+        int      *kib_peercredits;              /* # concurrent sends to 1 peer */
+
         struct ctl_table_header *kib_sysctl;    /* sysctl interface */
 } kib_tunables_t;
 
@@ -137,7 +170,7 @@ typedef struct
         __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 */
+        ib_mr_t          *ibp_handle;           /* mapped region handle */
         struct page      *ibp_pages[0];
 } kib_pages_t;
 
@@ -147,22 +180,19 @@ typedef struct
         __u64             kib_incarnation;      /* which one am I */
         int               kib_shutdown;         /* shut down? */
         atomic_t          kib_nthreads;         /* # live threads */
+        lnet_ni_t        *kib_ni;               /* _the_ openib interface */
 
         __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 */
+        int               kib_nonewpeers;       /* prevent new peers? */
         atomic_t          kib_npeers;           /* # peers extant */
         atomic_t          kib_nconns;           /* # connections extant */
 
@@ -174,6 +204,7 @@ typedef struct
         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 */
+        int               kib_connd_connecting; /* # connds connecting */
         spinlock_t        kib_connd_lock;       /* serialise */
 
         wait_queue_head_t kib_sched_waitq;      /* schedulers sleep here */
@@ -185,20 +216,19 @@ typedef struct
         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 */
 
+        int               kib_hca_idx;          /* my HCA number */
         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 */
+        ib_pd_t          *kib_pd;               /* protection domain */
 #if IBNAL_FMR
-        struct ib_fmr_pool *kib_fmr_pool;       /* fast memory region pool */
+        ib_fmr_pool_t    *kib_fmr_pool;         /* fast memory region pool */
 #endif
-        struct ib_cq     *kib_cq;               /* completion queue */
+        ib_cq_t          *kib_cq;               /* completion queue */
 
 } kib_data_t;
 
@@ -239,8 +269,8 @@ typedef struct kib_connparams
 typedef struct
 {
         union {
-                struct ib_mr    *mr;
-                struct ib_fmr   *fmr;
+                ib_mr_t         *mr;
+                ib_fmr_t        *fmr;
         }                 md_handle;
         __u32             md_lkey;
         __u32             md_rkey;
@@ -256,13 +286,13 @@ typedef struct
 
 typedef struct
 {
-        ptl_hdr_t         ibim_hdr;             /* portals header */
+        lnet_hdr_t        ibim_hdr;             /* portals header */
         char              ibim_payload[0];      /* piggy-backed payload */
 } WIRE_ATTR kib_immediate_msg_t;
 
 typedef struct
 {
-        ptl_hdr_t         ibrm_hdr;             /* portals header */
+        lnet_hdr_t        ibrm_hdr;             /* portals header */
         __u64             ibrm_cookie;          /* opaque completion cookie */
         kib_rdma_desc_t   ibrm_desc;            /* where to suck/blow */
 } WIRE_ATTR kib_rdma_msg_t;
@@ -296,8 +326,9 @@ typedef struct
         } WIRE_ATTR       ibm_u;
 } WIRE_ATTR kib_msg_t;
 
-#define IBNAL_MSG_MAGIC       0x0be91b91        /* unique magic */
-#define IBNAL_MSG_VERSION              2        /* current protocol version */
+#define IBNAL_MSG_MAGIC LNET_PROTO_OPENIB_MAGIC /* unique magic */
+#define IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD 2   /* previous protocol version */
+#define IBNAL_MSG_VERSION              3        /* current protocol version */
 
 #define IBNAL_MSG_SVCQRY            0xb0        /* service query */
 #define IBNAL_MSG_SVCRSP            0xb1        /* service response */
@@ -316,8 +347,7 @@ typedef struct kib_rx                           /* receive message */
 {
         struct list_head          rx_list;      /* queue for attention */
         struct kib_conn          *rx_conn;      /* owning conn */
-        int                       rx_rdma;      /* RDMA completion posted? */
-        int                       rx_posted;    /* posted? */
+        int                       rx_nob;       /* # bytes received (-1 while posted) */
         __u64                     rx_vaddr;     /* pre-mapped buffer (hca vaddr) */
         kib_msg_t                *rx_msg;       /* pre-mapped buffer (host vaddr) */
         struct ib_receive_param   rx_sp;        /* receive work item */
@@ -327,7 +357,6 @@ typedef struct kib_rx                           /* receive 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 kib_conn          *tx_conn;      /* owning conn */
         int                       tx_mapped;    /* mapped for RDMA? */
         int                       tx_sending;   /* # tx callbacks outstanding */
@@ -336,7 +365,7 @@ typedef struct kib_tx                           /* transmit message */
         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 */
+        lnet_msg_t               *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
         kib_md_t                  tx_md;        /* RDMA mapping (active/passive) */
         __u64                     tx_vaddr;     /* pre-mapped buffer (hca vaddr) */
         kib_msg_t                *tx_msg;       /* pre-mapped buffer (host vaddr) */
@@ -366,18 +395,22 @@ typedef struct kib_conn
         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 */
+        int                 ibc_version;        /* peer protocol version */
         atomic_t            ibc_refcount;       /* # users */
         int                 ibc_state;          /* what's happening */
-        atomic_t            ibc_nob;            /* # bytes buffered */
         int                 ibc_nsends_posted;  /* # uncompleted sends */
         int                 ibc_credits;        /* # credits I have */
         int                 ibc_outstanding_credits; /* # credits to return */
+        int                 ibc_reserved_credits; /* # credits for ACK/DONE msgs */
+        unsigned long       ibc_last_send;      /* time of last send */
+        struct list_head    ibc_tx_queue_nocred; /* sends that don't need a credit */
+        struct list_head    ibc_tx_queue_rsrvd; /* sends that need a reserved cred */
         struct list_head    ibc_tx_queue;       /* send queue */
         struct list_head    ibc_active_txs;     /* active tx awaiting completion */
         spinlock_t          ibc_lock;           /* serialise */
         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 */
+        ib_qp_t            *ibc_qp;             /* queue pair */
         __u32               ibc_qpn;            /* queue pair number */
         tTS_IB_CM_COMM_ID   ibc_comm_id;        /* connection ID? */
         kib_connreq_t      *ibc_connreq;        /* connection request state */
@@ -394,7 +427,7 @@ typedef struct kib_peer
 {
         struct list_head    ibp_list;           /* stash on global peer list */
         struct list_head    ibp_connd_list;     /* schedule on kib_connd_peers */
-        ptl_nid_t           ibp_nid;            /* who's on the other end(s) */
+        lnet_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 */
@@ -402,17 +435,69 @@ typedef struct kib_peer
         int                 ibp_persistence;    /* "known" peer refs */
         struct list_head    ibp_conns;          /* all active connections */
         struct list_head    ibp_tx_queue;       /* msgs waiting for a conn */
-        int                 ibp_connecting;     /* connecting+accepting */
+        int                 ibp_connecting;     /* current active connection attempts */
+        int                 ibp_accepting;      /* current passive connection attempts */
         unsigned long       ibp_reconnect_time; /* when reconnect may be attempted */
         unsigned long       ibp_reconnect_interval; /* exponential backoff */
+        int                 ibp_error;          /* errno on closing this peer */
+        cfs_time_t          ibp_last_alive;     /* when (in jiffies) I was last alive */
 } kib_peer_t;
 
-extern lib_nal_t       kibnal_lib;
 extern kib_data_t      kibnal_data;
 extern kib_tunables_t  kibnal_tunables;
 
+/******************************************************************************/
+
+/* these are purposely avoiding using local vars so they don't increase
+ * stack consumption. */
+
+#define kibnal_conn_addref(conn)                                \
+do {                                                            \
+        CDEBUG(D_NET, "conn[%p] (%d)++\n",                      \
+               (conn), atomic_read(&(conn)->ibc_refcount));     \
+        LASSERT(atomic_read(&(conn)->ibc_refcount) > 0);        \
+        atomic_inc(&(conn)->ibc_refcount);                      \
+} while (0)
+
+#define kibnal_conn_decref(conn)                                              \
+do {                                                                          \
+        unsigned long   flags;                                                \
+                                                                              \
+        CDEBUG(D_NET, "conn[%p] (%d)--\n",                                    \
+               (conn), atomic_read(&(conn)->ibc_refcount));                   \
+        LASSERT(atomic_read(&(conn)->ibc_refcount) > 0);                      \
+        if (atomic_dec_and_test(&(conn)->ibc_refcount)) {                     \
+                spin_lock_irqsave(&kibnal_data.kib_reaper_lock, flags);       \
+                list_add_tail(&(conn)->ibc_list,                              \
+                              &kibnal_data.kib_reaper_conns);                 \
+                wake_up(&kibnal_data.kib_reaper_waitq);                       \
+                spin_unlock_irqrestore(&kibnal_data.kib_reaper_lock, flags);  \
+        }                                                                     \
+} while (0)
+
+#define kibnal_peer_addref(peer)                                \
+do {                                                            \
+        CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n",                \
+               (peer), libcfs_nid2str((peer)->ibp_nid),         \
+               atomic_read (&(peer)->ibp_refcount));            \
+        LASSERT(atomic_read(&(peer)->ibp_refcount) > 0);        \
+        atomic_inc(&(peer)->ibp_refcount);                      \
+} while (0)
+
+#define kibnal_peer_decref(peer)                                \
+do {                                                            \
+        CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n",                \
+               (peer), libcfs_nid2str((peer)->ibp_nid),         \
+               atomic_read (&(peer)->ibp_refcount));            \
+        LASSERT(atomic_read(&(peer)->ibp_refcount) > 0);        \
+        if (atomic_dec_and_test(&(peer)->ibp_refcount))         \
+                kibnal_destroy_peer(peer);                      \
+} while (0)
+
+/******************************************************************************/
+
 static inline struct list_head *
-kibnal_nid2peerlist (ptl_nid_t nid)
+kibnal_nid2peerlist (lnet_nid_t nid)
 {
         unsigned int hash = ((unsigned int)nid) % kibnal_data.kib_peer_hash_size;
 
@@ -429,42 +514,57 @@ kibnal_peer_active(kib_peer_t *peer)
 static inline void
 kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
 {
-        /* CAVEAT EMPTOR: tx takes caller's ref on conn */
+        struct list_head      *q;
 
         LASSERT (tx->tx_nsp > 0);               /* work items set up */
         LASSERT (tx->tx_conn == NULL);          /* only set here */
 
+        kibnal_conn_addref(conn);
         tx->tx_conn = conn;
-        tx->tx_deadline = jiffies + kibnal_tunables.kib_io_timeout * HZ;
-        list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
-}
-
-#if 0
-static inline void
-kibnal_show_rdma_attr (kib_conn_t *conn)
-{
-        struct ib_qp_attribute qp_attr;
-        int                    rc;
-
-        memset (&qp_attr, 0, sizeof(qp_attr));
-        rc = ib_qp_query(conn->ibc_qp, &qp_attr);
-        if (rc != 0) {
-                CERROR ("Can't get qp attrs: %d\n", rc);
-                return;
+        tx->tx_deadline = jiffies + *kibnal_tunables.kib_timeout * HZ;
+
+        if (conn->ibc_version == IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD) {
+                /* All messages have simple credit control */
+                q = &conn->ibc_tx_queue;
+        } else {
+                LASSERT (conn->ibc_version == IBNAL_MSG_VERSION);
+                
+                switch (tx->tx_msg->ibm_type) {
+                case IBNAL_MSG_PUT_RDMA:
+                case IBNAL_MSG_GET_RDMA:
+                        /* RDMA request: reserve a buffer for the RDMA reply
+                         * before sending */
+                        q = &conn->ibc_tx_queue_rsrvd;
+                        break;
+
+                case IBNAL_MSG_PUT_DONE:
+                case IBNAL_MSG_GET_DONE:
+                        /* RDMA completion: no credits; peer has reserved a
+                         * reply buffer */
+                        q = &conn->ibc_tx_queue_nocred;
+                        break;
+                
+                case IBNAL_MSG_NOOP:
+                case IBNAL_MSG_IMMEDIATE:
+                        /* Otherwise: consume a credit before sending */
+                        q = &conn->ibc_tx_queue;
+                        break;
+                
+                default:
+                        LBUG();
+                        q = NULL;
+                }
         }
 
-        CWARN ("RDMA CAPABILITY: write %s read %s\n",
-               (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ?
-               (qp_attr.enable_rdma_write ? "enabled" : "disabled") : "invalid",
-               (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ?
-               (qp_attr.enable_rdma_read ? "enabled" : "disabled") : "invalid");
+        list_add_tail(&tx->tx_list, q);
 }
-#endif
 
-static inline __u64
-kibnal_page2phys (struct page *p)
+static inline int
+kibnal_send_keepalive(kib_conn_t *conn) 
 {
-        return page_to_phys(p);
+        return (*kibnal_tunables.kib_keepalive > 0) &&
+                time_after(jiffies, conn->ibc_last_send +
+                           *kibnal_tunables.kib_keepalive*HZ);
 }
 
 /* CAVEAT EMPTOR:
@@ -494,38 +594,63 @@ 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
+#if (IB_NTXRXPARAMS == 3)
+static inline int
+kibnal_ib_send(ib_qp_t *qp, struct ib_send_param *p)
+{
+        return ib_send(qp, p, 1);
+}
+
+static inline int
+kibnal_ib_receive(ib_qp_t *qp, struct ib_receive_param *p)
+{
+        return ib_receive(qp, p, 1);
+}
+#elif (IB_NTXRXPARAMS == 4)
+static inline int
+kibnal_ib_send(ib_qp_t *qp, struct ib_send_param *p)
+{
+        return ib_send(qp, p, 1, NULL);
+}
+
+static inline int
+kibnal_ib_receive(ib_qp_t *qp, struct ib_receive_param *p)
+{
+        return ib_receive(qp, p, 1, NULL);
+}
+#else
+ #error "IB_NTXRXPARAMS not set correctly"
 #endif
 
+int kibnal_startup (lnet_ni_t *ni);
+void kibnal_shutdown (lnet_ni_t *ni);
+int kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
+int kibnal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
+int kibnal_eager_recv (lnet_ni_t *ni, void *private, 
+                       lnet_msg_t *lntmsg, void **new_private);
+int kibnal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, 
+                int delayed, unsigned int niov, 
+                struct iovec *iov, lnet_kiov_t *kiov,
+                unsigned int offset, unsigned int mlen, unsigned int rlen);
+int kibnal_accept(lnet_ni_t *ni, struct socket *sock);
+
 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_pack_msg(kib_msg_t *msg, int version, int credits, 
+                            lnet_nid_t dstnid, __u64 dststamp);
+extern int kibnal_unpack_msg(kib_msg_t *msg, int expected_version, 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 int kibnal_create_peer (kib_peer_t **peerp, lnet_nid_t nid);
+extern void kibnal_destroy_peer (kib_peer_t *peer);
+extern int kibnal_add_persistent_peer(lnet_nid_t nid, __u32 ip, int port);
+extern int kibnal_del_peer (lnet_nid_t nid);
+extern kib_peer_t *kibnal_find_peer_locked (lnet_nid_t nid);
 extern void kibnal_unlink_peer_locked (kib_peer_t *peer);
+extern void kibnal_peer_alive(kib_peer_t *peer);
 extern int  kibnal_close_stale_conns_locked (kib_peer_t *peer,
                                               __u64 incarnation);
 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);
@@ -548,16 +673,15 @@ 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_callback (ib_cq_t *cq, struct ib_cq_entry *e, void *arg);
+extern void kibnal_txlist_done (struct list_head *txlist, int status);
 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,
+                                      kib_rx_t *rx, lnet_msg_t *lntmsg,
                                       unsigned int niov,
-                                      struct iovec *iov, ptl_kiov_t *kiov,
+                                      struct iovec *iov, lnet_kiov_t *kiov,
                                       int offset, int nob);
 
-
-
-
-
+extern int  kibnal_tunables_init(void);
+extern void kibnal_tunables_fini(void);