lnet/klnds/iiblnd/iiblnd.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004 Cluster File Systems, Inc.
   5  *   Author: Eric Barton <eric@bartonsoftware.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27
  28 #include <linux/config.h>
  29 #include <linux/module.h>
  30 #include <linux/kernel.h>
  31 #include <linux/mm.h>
  32 #include <linux/string.h>
  33 #include <linux/stat.h>
  34 #include <linux/errno.h>
  35 #include <linux/smp_lock.h>
  36 #include <linux/unistd.h>
  37 #include <linux/uio.h>
  38
  39 #include <asm/system.h>
  40 #include <asm/uaccess.h>
  41 #include <asm/io.h>
  42
  43 #include <linux/init.h>
  44 #include <linux/fs.h>
  45 #include <linux/file.h>
  46 #include <linux/stat.h>
  47 #include <linux/list.h>
  48 #include <linux/kmod.h>
  49 #include <linux/sysctl.h>
  50
  51 #define DEBUG_SUBSYSTEM S_LND
  52
  53 #include <libcfs/kp30.h>
  54 #include <lnet/lnet.h>
  55 #include <lnet/lib-lnet.h>
  56
  57 #include <linux/iba/ibt.h>
  58
  59 #define GCC_VERSION (__GNUC__ * 10000 \
  60                 + __GNUC_MINOR__ * 100 \
  61                 + __GNUC_PATCHLEVEL__)
  62
  63 /* Test for GCC > 3.2.2 */
  64 #if GCC_VERSION <= 30202
  65 /* GCC 3.2.2, and presumably several versions before it, will
  66  * miscompile this driver. See
  67  * http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9853. */
  68 #error Invalid GCC version. Must use GCC >= 3.2.3
  69 #endif
  70
  71 #ifdef CONFIG_SMP
  72 # define IBNAL_N_SCHED      num_online_cpus()   /* # schedulers */
  73 #else
  74 # define IBNAL_N_SCHED      1                   /* # schedulers */
  75 #endif
  76
  77 #define IBNAL_USE_FMR                0          /* map on demand v. use whole mem mapping */
  78 #define KIBLND_DETAILED_DEBUG        0
  79
  80 /* tunables fixed at compile time */
  81 #define IBNAL_PEER_HASH_SIZE         101        /* # peer lists */
  82 #define IBNAL_RESCHED                100        /* # scheduler loops before reschedule */
  83 #define IBNAL_MSG_QUEUE_SIZE         8          /* # messages/RDMAs in-flight */
  84 #define IBNAL_CREDIT_HIGHWATER       7          /* when to eagerly return credits */
  85 #define IBNAL_MSG_SIZE              (4<<10)     /* max size of queued messages (inc hdr) */
  86 #define IBNAL_RDMA_BASE              0x0eeb0000
  87 #define IBNAL_STARTING_PSN           1
  88
  89 /* QP tunables */
  90 /* 7 indicates infinite retry attempts, Infinicon recommended 5 */
  91 #define IBNAL_RETRY                  5          /* # times to retry */
  92 #define IBNAL_RNR_RETRY              5          /*  */
  93 #define IBNAL_CM_RETRY               5          /* # times to retry connection */
  94 #define IBNAL_FLOW_CONTROL           1
  95 #define IBNAL_ACK_TIMEOUT            20         /* supposedly 4 secs */
  96 #define IBNAL_EE_FLOW                1
  97 #define IBNAL_LOCAL_SUB              1
  98 #define IBNAL_FAILOVER_ACCEPTED      0
  99
 100 /************************/
 101 /* derived constants... */
 102
 103 /* TX messages (shared by all connections) */
 104 #define IBNAL_TX_MSGS()       (*kibnal_tunables.kib_ntx)
 105 #define IBNAL_TX_MSG_BYTES()  (IBNAL_TX_MSGS() * IBNAL_MSG_SIZE)
 106 #define IBNAL_TX_MSG_PAGES()  ((IBNAL_TX_MSG_BYTES() + PAGE_SIZE - 1)/PAGE_SIZE)
 107
 108 #if IBNAL_USE_FMR
 109 # define IBNAL_MAX_RDMA_FRAGS 1
 110 # define IBNAL_CONCURRENT_SENDS IBNAL_RX_MSGS
 111 #else
 112 # define IBNAL_MAX_RDMA_FRAGS LNET_MAX_IOV
 113 # define IBNAL_CONCURRENT_SENDS IBNAL_MSG_QUEUE_SIZE
 114 #endif
 115
 116 /* RX messages (per connection) */
 117 #define IBNAL_RX_MSGS         (IBNAL_MSG_QUEUE_SIZE * 2)
 118 #define IBNAL_RX_MSG_BYTES    (IBNAL_RX_MSGS * IBNAL_MSG_SIZE)
 119 #define IBNAL_RX_MSG_PAGES    ((IBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE)
 120
 121 #define IBNAL_CQ_ENTRIES()  (IBNAL_TX_MSGS() * (1 + IBNAL_MAX_RDMA_FRAGS) +             \
 122                              (IBNAL_RX_MSGS * *kibnal_tunables.kib_concurrent_peers))
 123
 124 typedef struct
 125 {
 126         char            **kib_hca_basename;     /* HCA base name */
 127         char            **kib_ipif_basename;    /* IPoIB interface base name */
 128         char            **kib_service_name;     /* global service name */
 129         unsigned int     *kib_service_number;   /* global service number */
 130         int              *kib_min_reconnect_interval; /* min connect retry seconds... */
 131         int              *kib_max_reconnect_interval; /* max connect retry seconds */
 132         int              *kib_concurrent_peers; /* max # peers */
 133         int              *kib_cksum;            /* checksum kib_msg_t? */
 134         int              *kib_timeout;          /* comms timeout (seconds) */
 135         int              *kib_keepalive;        /* keepalive timeout (seconds) */
 136         int              *kib_ntx;              /* # tx descs */
 137         int              *kib_credits;          /* # concurrent sends */
 138         int              *kib_peercredits;      /* # concurrent sends to 1 peer */
 139         int              *kib_sd_retries;       /* # concurrent sends to 1 peer */
 140         int              *kib_concurrent_sends; /* send work queue sizing */
 141 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
 142         struct ctl_table_header *kib_sysctl;    /* sysctl interface */
 143 #endif
 144 } kib_tunables_t;
 145
 146 /* NB The Infinicon stack has specific typedefs for some things
 147  * (e.g. IB_{L,R}_KEY), that just map back to __u32 etc */
 148 typedef struct
 149 {
 150         int               ibp_npages;           /* # pages */
 151         struct page      *ibp_pages[0];
 152 } kib_pages_t;
 153
 154 typedef struct
 155 {
 156         IB_HANDLE         md_handle;
 157         __u32             md_lkey;
 158         __u32             md_rkey;
 159         __u64             md_addr;
 160 } kib_md_t;
 161
 162 typedef struct
 163 {
 164         int               kib_init;             /* initialisation state */
 165         __u64             kib_incarnation;      /* which one am I */
 166         int               kib_shutdown;         /* shut down? */
 167         atomic_t          kib_nthreads;         /* # live threads */
 168         lnet_ni_t        *kib_ni;               /* _the_ iib instance */
 169
 170         __u64             kib_port_guid;        /* my GUID (lo 64 of GID)*/
 171         __u16             kib_port_pkey;        /* my pkey, whatever that is */
 172         struct semaphore  kib_listener_signal;  /* signal completion */
 173         IB_HANDLE         kib_listener_cep;     /* connection end point */
 174
 175         rwlock_t          kib_global_lock;      /* stabilize peer/conn ops */
 176         int               kib_ready;            /* CQ callback fired */
 177         int               kib_checking_cq;      /* a scheduler is checking the CQ */
 178
 179         struct list_head *kib_peers;            /* hash table of all my known peers */
 180         int               kib_peer_hash_size;   /* size of kib_peers */
 181         atomic_t          kib_npeers;           /* # peers extant */
 182         atomic_t          kib_nconns;           /* # connections extant */
 183
 184         struct list_head  kib_connd_zombies;    /* connections to free */
 185         struct list_head  kib_connd_conns;      /* connections to progress */
 186         struct list_head  kib_connd_peers;      /* peers waiting for a connection */
 187         wait_queue_head_t kib_connd_waitq;      /* connection daemon sleep here */
 188         spinlock_t        kib_connd_lock;       /* serialise */
 189
 190         wait_queue_head_t kib_sched_waitq;      /* schedulers sleep here */
 191         spinlock_t        kib_sched_lock;       /* serialise */
 192
 193         struct kib_tx    *kib_tx_descs;         /* all the tx descriptors */
 194         kib_pages_t      *kib_tx_pages;         /* premapped tx msg pages */
 195
 196         struct list_head  kib_idle_txs;         /* idle tx descriptors */
 197         __u64             kib_next_tx_cookie;   /* RDMA completion cookie */
 198         spinlock_t        kib_tx_lock;          /* serialise */
 199
 200         IB_HANDLE         kib_hca;              /* The HCA */
 201         int               kib_port;             /* port on the device */
 202         IB_HANDLE         kib_pd;               /* protection domain */
 203         IB_HANDLE         kib_sd;               /* SD handle */
 204         IB_HANDLE         kib_cq;               /* completion queue */
 205         kib_md_t          kib_whole_mem;        /* whole-mem registration */
 206
 207         int               kib_hca_idx;          /* my HCA number */
 208         uint64            kib_hca_guids[8];     /* all the HCA guids */
 209         IB_CA_ATTRIBUTES  kib_hca_attrs;        /* where to get HCA attrs */
 210
 211         COMMAND_CONTROL_PARAMETERS kib_sdretry; /* control SD query retries */
 212 } kib_data_t;
 213
 214 #define IBNAL_INIT_NOTHING         0
 215 #define IBNAL_INIT_DATA            1
 216 #define IBNAL_INIT_LIB             2
 217 #define IBNAL_INIT_HCA             3
 218 #define IBNAL_INIT_PORTATTRS       4
 219 #define IBNAL_INIT_SD              5
 220 #define IBNAL_INIT_PD              6
 221 #define IBNAL_INIT_MD              7
 222 #define IBNAL_INIT_TXD             8
 223 #define IBNAL_INIT_CQ              9
 224 #define IBNAL_INIT_ALL             10
 225
 226 /************************************************************************
 227  * Wire message structs.
 228  * These are sent in sender's byte order (i.e. receiver flips).
 229  * CAVEAT EMPTOR: other structs communicated between nodes (e.g. MAD
 230  * private data and SM service info), is LE on the wire.
 231  */
 232
 233 typedef struct kib_connparams
 234 {
 235         __u32             ibcp_queue_depth;
 236         __u32             ibcp_max_msg_size;
 237         __u32             ibcp_max_frags;
 238 } WIRE_ATTR kib_connparams_t;
 239
 240 typedef struct
 241 {
 242         lnet_hdr_t        ibim_hdr;             /* portals header */
 243         char              ibim_payload[0];      /* piggy-backed payload */
 244 } WIRE_ATTR kib_immediate_msg_t;
 245
 246 #if IBNAL_USE_FMR
 247 typedef struct
 248 {
 249         __u64             rd_addr;              /* IO VMA address */
 250         __u32             rd_nob;               /* # of bytes */
 251         __u32             rd_key;               /* remote key */
 252 } WIRE_ATTR kib_rdma_desc_t;
 253 #else
 254 typedef struct
 255 {
 256         __u32             rf_nob;               /* # of bytes */
 257         __u64             rf_addr;              /* remote io vaddr */
 258 } WIRE_ATTR kib_rdma_frag_t;
 259
 260 typedef struct
 261 {
 262         __u32             rd_key;               /* local/remote key */
 263         __u32             rd_nfrag;             /* # fragments */
 264         kib_rdma_frag_t   rd_frags[0];          /* buffer frags */
 265 } WIRE_ATTR kib_rdma_desc_t;
 266 #endif
 267
 268 typedef struct
 269 {
 270         lnet_hdr_t        ibprm_hdr;            /* LNET header */
 271         __u64             ibprm_cookie;         /* opaque completion cookie */
 272 } WIRE_ATTR kib_putreq_msg_t;
 273
 274 typedef struct
 275 {
 276         __u64             ibpam_src_cookie;     /* reflected completion cookie */
 277         __u64             ibpam_dst_cookie;     /* opaque completion cookie */
 278         kib_rdma_desc_t   ibpam_rd;             /* sender's sink buffer */
 279 } WIRE_ATTR kib_putack_msg_t;
 280
 281 typedef struct
 282 {
 283         lnet_hdr_t        ibgm_hdr;             /* LNET header */
 284         __u64             ibgm_cookie;          /* opaque completion cookie */
 285         kib_rdma_desc_t   ibgm_rd;              /* sender's sink buffer */
 286 } WIRE_ATTR kib_get_msg_t;
 287
 288 typedef struct
 289 {
 290         __u64             ibcm_cookie;          /* opaque completion cookie */
 291         __u32             ibcm_status;          /* completion status */
 292 } WIRE_ATTR kib_completion_msg_t;
 293
 294 typedef struct
 295 {
 296         /* First 2 fields fixed FOR ALL TIME */
 297         __u32             ibm_magic;            /* I'm an openibnal message */
 298         __u16             ibm_version;          /* this is my version number */
 299
 300         __u8              ibm_type;             /* msg type */
 301         __u8              ibm_credits;          /* returned credits */
 302         __u32             ibm_nob;              /* # bytes in whole message */
 303         __u32             ibm_cksum;            /* checksum (0 == no checksum) */
 304         __u64             ibm_srcnid;           /* sender's NID */
 305         __u64             ibm_srcstamp;         /* sender's incarnation */
 306         __u64             ibm_dstnid;           /* destination's NID */
 307         __u64             ibm_dststamp;         /* destination's incarnation */
 308         __u64             ibm_seq;              /* sequence number */
 309
 310         union {
 311                 kib_connparams_t      connparams;
 312                 kib_immediate_msg_t   immediate;
 313                 kib_putreq_msg_t      putreq;
 314                 kib_putack_msg_t      putack;
 315                 kib_get_msg_t         get;
 316                 kib_completion_msg_t  completion;
 317         } WIRE_ATTR ibm_u;
 318 } WIRE_ATTR kib_msg_t;
 319
 320 #define IBNAL_MSG_MAGIC LNET_PROTO_IIB_MAGIC    /* unique magic */
 321 #define IBNAL_MSG_VERSION              2        /* current protocol version */
 322 #define IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD 1   /* previous version */
 323
 324 #define IBNAL_MSG_CONNREQ           0xc0        /* connection request */
 325 #define IBNAL_MSG_CONNACK           0xc1        /* connection acknowledge */
 326 #define IBNAL_MSG_NOOP              0xd0        /* nothing (just credits) */
 327 #define IBNAL_MSG_IMMEDIATE         0xd1        /* immediate */
 328 #define IBNAL_MSG_PUT_REQ           0xd2        /* putreq (src->sink) */
 329 #define IBNAL_MSG_PUT_NAK           0xd3        /* completion (sink->src) */
 330 #define IBNAL_MSG_PUT_ACK           0xd4        /* putack (sink->src) */
 331 #define IBNAL_MSG_PUT_DONE          0xd5        /* completion (src->sink) */
 332 #define IBNAL_MSG_GET_REQ           0xd6        /* getreq (sink->src) */
 333 #define IBNAL_MSG_GET_DONE          0xd7        /* completion (src->sink: all OK) */
 334
 335 /* connection rejection reasons */
 336 #define IBNAL_REJECT_CONN_RACE       0          /* You lost connection race */
 337 #define IBNAL_REJECT_NO_RESOURCES    1          /* Out of memory/conns etc */
 338 #define IBNAL_REJECT_FATAL           2          /* Anything else */
 339
 340 /***********************************************************************/
 341
 342 typedef struct kib_rx                           /* receive message */
 343 {
 344         struct list_head          rx_list;      /* queue for attention */
 345         struct kib_conn          *rx_conn;      /* owning conn */
 346         int                       rx_nob;       /* # bytes received (-1 while posted) */
 347         __u64                     rx_hca_msg;   /* pre-mapped buffer (hca vaddr) */
 348         kib_msg_t                *rx_msg;       /* pre-mapped buffer (host vaddr) */
 349         IB_WORK_REQ2              rx_wrq;
 350         IB_LOCAL_DATASEGMENT      rx_gl;        /* and its memory */
 351 } kib_rx_t;
 352
 353 typedef struct kib_tx                           /* transmit message */
 354 {
 355         struct list_head          tx_list;      /* queue on idle_txs ibc_tx_queue etc. */
 356         struct kib_conn          *tx_conn;      /* owning conn */
 357         int                       tx_mapped;    /* mapped for RDMA? */
 358         int                       tx_sending;   /* # tx callbacks outstanding */
 359         int                       tx_queued;    /* queued for sending */
 360         int                       tx_waiting;   /* waiting for peer */
 361         int                       tx_status;    /* completion status */
 362         unsigned long             tx_deadline;  /* completion deadline */
 363         __u64                     tx_cookie;    /* completion cookie */
 364         lnet_msg_t               *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
 365         kib_msg_t                *tx_msg;       /* pre-mapped buffer (host vaddr) */
 366         __u64                     tx_hca_msg;   /* pre-mapped buffer (HCA vaddr) */
 367         int                       tx_nwrq;      /* # send work items */
 368 #if IBNAL_USE_FMR
 369         IB_WORK_REQ2              tx_wrq[2];    /* send work items... */
 370         IB_LOCAL_DATASEGMENT      tx_gl[2];     /* ...and their memory */
 371         kib_rdma_desc_t           tx_rd[1];     /* rdma descriptor */
 372         kib_md_t                  tx_md;        /* mapping */
 373         __u64                    *tx_pages;     /* page phys addrs */
 374 #else
 375         IB_WORK_REQ2             *tx_wrq;       /* send work items... */
 376         IB_LOCAL_DATASEGMENT     *tx_gl;        /* ...and their memory */
 377         kib_rdma_desc_t          *tx_rd;        /* rdma descriptor (src buffers) */
 378 #endif
 379 } kib_tx_t;
 380
 381 typedef struct
 382 {
 383         /* scratchpad during connection establishment */
 384         IB_QP_ATTRIBUTES_QUERY cv_qpattrs;
 385         QUERY                  cv_query;
 386         IB_SERVICE_RECORD      cv_svcrec;
 387         IB_PATH_RECORD         cv_path;
 388         CM_CONN_INFO           cv_cmci;
 389 } kib_connvars_t;
 390
 391 typedef struct kib_conn
 392 {
 393         struct kib_peer    *ibc_peer;           /* owning peer */
 394         struct list_head    ibc_list;           /* stash on peer's conn list */
 395         __u64               ibc_incarnation;    /* which instance of the peer */
 396         __u64               ibc_txseq;          /* tx sequence number */
 397         __u64               ibc_rxseq;          /* rx sequence number */
 398         __u32               ibc_version;        /* peer protocol version */
 399         atomic_t            ibc_refcount;       /* # users */
 400         int                 ibc_state;          /* what's happening */
 401         int                 ibc_nsends_posted;  /* # uncompleted sends */
 402         int                 ibc_credits;        /* # credits I have */
 403         int                 ibc_outstanding_credits; /* # credits to return */
 404         int                 ibc_reserved_credits; /* # credits for ACK/DONE msgs */
 405         unsigned long       ibc_last_send;      /* time of last send */
 406         struct list_head    ibc_early_rxs;      /* rxs completed before ESTABLISHED */
 407         struct list_head    ibc_tx_queue_nocred; /* sends that don't need a cred */
 408         struct list_head    ibc_tx_queue_rsrvd; /* sends that need a reserved cred */
 409         struct list_head    ibc_tx_queue;       /* send queue */
 410         struct list_head    ibc_active_txs;     /* active tx awaiting completion */
 411         spinlock_t          ibc_lock;           /* serialise */
 412         kib_rx_t           *ibc_rxs;            /* the rx descs */
 413         kib_pages_t        *ibc_rx_pages;       /* premapped rx msg pages */
 414         IB_HANDLE           ibc_qp;             /* queue pair */
 415         IB_HANDLE           ibc_cep;            /* CM endpoint */
 416         kib_connvars_t     *ibc_cvars;          /* connection scratchpad */
 417 } kib_conn_t;
 418
 419 #define IBNAL_CONN_INIT_NOTHING      0          /* initial state */
 420 #define IBNAL_CONN_INIT_QP           1          /* ibc_qp set up */
 421 #define IBNAL_CONN_CONNECTING        2          /* started to connect */
 422 #define IBNAL_CONN_ESTABLISHED       3          /* connection established */
 423 #define IBNAL_CONN_DISCONNECTING     4          /* to send disconnect req */
 424 #define IBNAL_CONN_DISCONNECTED      5          /* no more QP or CM traffic */
 425
 426 /* types of connection */
 427 #define IBNAL_CONN_ACTIVE            0          /* active connect */
 428 #define IBNAL_CONN_PASSIVE           1          /* passive connect */
 429 #define IBNAL_CONN_WAITING           2          /* waiting for connect */
 430
 431 typedef struct kib_peer
 432 {
 433         struct list_head    ibp_list;           /* stash on global peer list */
 434         struct list_head    ibp_connd_list;     /* schedule on kib_connd_peers */
 435         lnet_nid_t          ibp_nid;            /* who's on the other end(s) */
 436         atomic_t            ibp_refcount;       /* # users */
 437         int                 ibp_persistence;    /* "known" peer refs */
 438         int                 ibp_version;        /* protocol version */
 439         struct list_head    ibp_conns;          /* all active connections */
 440         struct list_head    ibp_tx_queue;       /* msgs waiting for a conn */
 441         int                 ibp_connecting;     /* active connects in progress */
 442         int                 ibp_accepting;      /* passive connects in progress */
 443         int                 ibp_passivewait;    /* waiting for peer to connect */
 444         unsigned long       ibp_passivewait_deadline; /* when passive wait must complete */
 445         unsigned long       ibp_reconnect_time; /* when reconnect may be attempted */
 446         unsigned long       ibp_reconnect_interval; /* exponential backoff */
 447         int                 ibp_error;          /* errno on closing this peer */
 448         cfs_time_t          ibp_last_alive;     /* when (in jiffies) I was last alive */
 449 } kib_peer_t;
 450
 451
 452 extern kib_data_t      kibnal_data;
 453 extern kib_tunables_t  kibnal_tunables;
 454
 455 /******************************************************************************/
 456
 457 /* these are purposely avoiding using local vars so they don't increase
 458  * stack consumption. */
 459
 460 #define kibnal_conn_addref(conn)                                \
 461 do {                                                            \
 462         CDEBUG(D_NET, "conn[%p] (%d)++\n",                      \
 463                (conn), atomic_read(&(conn)->ibc_refcount));     \
 464         LASSERT(atomic_read(&(conn)->ibc_refcount) > 0);        \
 465         atomic_inc(&(conn)->ibc_refcount);                      \
 466 } while (0)
 467
 468 #define kibnal_conn_decref(conn)                                              \
 469 do {                                                                          \
 470         unsigned long   flags;                                                \
 471                                                                               \
 472         CDEBUG(D_NET, "conn[%p] (%d)--\n",                                    \
 473                (conn), atomic_read(&(conn)->ibc_refcount));                   \
 474         LASSERT(atomic_read(&(conn)->ibc_refcount) > 0);                      \
 475         if (atomic_dec_and_test(&(conn)->ibc_refcount)) {                     \
 476                 spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags);        \
 477                 list_add_tail(&(conn)->ibc_list,                              \
 478                               &kibnal_data.kib_connd_zombies);                \
 479                 wake_up(&kibnal_data.kib_connd_waitq);                        \
 480                 spin_unlock_irqrestore(&kibnal_data.kib_connd_lock, flags);   \
 481         }                                                                     \
 482 } while (0)
 483
 484 #define kibnal_peer_addref(peer)                                \
 485 do {                                                            \
 486         CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n",                \
 487                (peer), libcfs_nid2str((peer)->ibp_nid),         \
 488                atomic_read (&(peer)->ibp_refcount));            \
 489         LASSERT(atomic_read(&(peer)->ibp_refcount) > 0);        \
 490         atomic_inc(&(peer)->ibp_refcount);                      \
 491 } while (0)
 492
 493 #define kibnal_peer_decref(peer)                                \
 494 do {                                                            \
 495         CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n",                \
 496                (peer), libcfs_nid2str((peer)->ibp_nid),         \
 497                atomic_read (&(peer)->ibp_refcount));            \
 498         LASSERT(atomic_read(&(peer)->ibp_refcount) > 0);        \
 499         if (atomic_dec_and_test(&(peer)->ibp_refcount))         \
 500                 kibnal_destroy_peer(peer);                      \
 501 } while (0)
 502
 503 /******************************************************************************/
 504
 505 static inline struct list_head *
 506 kibnal_nid2peerlist (lnet_nid_t nid)
 507 {
 508         unsigned int hash = ((unsigned int)nid) % kibnal_data.kib_peer_hash_size;
 509
 510         return (&kibnal_data.kib_peers [hash]);
 511 }
 512
 513 static inline int
 514 kibnal_peer_active(kib_peer_t *peer)
 515 {
 516         /* Am I in the peer hash table? */
 517         return (!list_empty(&peer->ibp_list));
 518 }
 519
 520 static inline int
 521 kibnal_peer_connecting(kib_peer_t *peer)
 522 {
 523         /* Am I expecting a connection to materialise? */
 524         return (peer->ibp_connecting != 0 ||
 525                 peer->ibp_accepting != 0 ||
 526                 peer->ibp_passivewait);
 527 }
 528
 529 static inline void
 530 kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
 531 {
 532         struct list_head  *q;
 533
 534         LASSERT (tx->tx_nwrq > 0);              /* work items set up */
 535         LASSERT (!tx->tx_queued);               /* not queued for sending already */
 536
 537         tx->tx_queued = 1;
 538         tx->tx_deadline = jiffies + (*kibnal_tunables.kib_timeout * HZ);
 539
 540         if (tx->tx_conn == NULL) {
 541                 kibnal_conn_addref(conn);
 542                 tx->tx_conn = conn;
 543                 LASSERT (tx->tx_msg->ibm_type != IBNAL_MSG_PUT_DONE);
 544         } else {
 545                 LASSERT (tx->tx_conn == conn);
 546                 LASSERT (tx->tx_msg->ibm_type == IBNAL_MSG_PUT_DONE);
 547         }
 548
 549         if (conn->ibc_version == IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD) {
 550                 /* All messages have simple credit control */
 551                 q = &conn->ibc_tx_queue;
 552         } else {
 553                 LASSERT (conn->ibc_version == IBNAL_MSG_VERSION);
 554
 555                 switch (tx->tx_msg->ibm_type) {
 556                 case IBNAL_MSG_PUT_REQ:
 557                 case IBNAL_MSG_GET_REQ:
 558                         /* RDMA request: reserve a buffer for the RDMA reply
 559                          * before sending */
 560                         q = &conn->ibc_tx_queue_rsrvd;
 561                         break;
 562
 563                 case IBNAL_MSG_PUT_NAK:
 564                 case IBNAL_MSG_PUT_ACK:
 565                 case IBNAL_MSG_PUT_DONE:
 566                 case IBNAL_MSG_GET_DONE:
 567                         /* RDMA reply/completion: no credits; peer has reserved
 568                          * a reply buffer */
 569                         q = &conn->ibc_tx_queue_nocred;
 570                         break;
 571
 572                 case IBNAL_MSG_NOOP:
 573                 case IBNAL_MSG_IMMEDIATE:
 574                         /* Otherwise: consume a credit before sending */
 575                         q = &conn->ibc_tx_queue;
 576                         break;
 577
 578                 default:
 579                         LBUG();
 580                         q = NULL;
 581                 }
 582         }
 583
 584         list_add_tail(&tx->tx_list, q);
 585 }
 586
 587 static inline int
 588 kibnal_send_keepalive(kib_conn_t *conn)
 589 {
 590         return (*kibnal_tunables.kib_keepalive > 0) &&
 591                 time_after(jiffies, conn->ibc_last_send +
 592                            *kibnal_tunables.kib_keepalive*HZ);
 593 }
 594
 595 #define KIBNAL_SERVICE_KEY_MASK  (IB_SERVICE_RECORD_COMP_SERVICENAME |          \
 596                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_1 |       \
 597                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_2 |       \
 598                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_3 |       \
 599                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_4 |       \
 600                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_5 |       \
 601                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_6 |       \
 602                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_7 |       \
 603                                   IB_SERVICE_RECORD_COMP_SERVICEDATA8_8)
 604
 605 static inline __u64*
 606 kibnal_service_nid_field(IB_SERVICE_RECORD *srv)
 607 {
 608         /* must be consistent with KIBNAL_SERVICE_KEY_MASK */
 609         return (__u64 *)srv->ServiceData8;
 610 }
 611
 612 static inline void
 613 kibnal_set_service_keys(IB_SERVICE_RECORD *srv, lnet_nid_t nid)
 614 {
 615         char *svc_name = *kibnal_tunables.kib_service_name;
 616
 617         LASSERT (strlen(svc_name) < sizeof(srv->ServiceName));
 618         memset (srv->ServiceName, 0, sizeof(srv->ServiceName));
 619         strcpy (srv->ServiceName, svc_name);
 620
 621         *kibnal_service_nid_field(srv) = cpu_to_le64(nid);
 622 }
 623
 624 /* CAVEAT EMPTOR: We rely on tx/rx descriptor alignment to allow us to use the
 625  * lowest 2 bits of the work request id to stash the work item type (the op
 626  * field is not valid when the wc completes in error). */
 627
 628 #define IBNAL_WID_TX    0
 629 #define IBNAL_WID_RX    1
 630 #define IBNAL_WID_RDMA  2
 631 #define IBNAL_WID_MASK  3UL
 632
 633 static inline __u64
 634 kibnal_ptr2wreqid (void *ptr, int type)
 635 {
 636         unsigned long lptr = (unsigned long)ptr;
 637
 638         LASSERT ((lptr & IBNAL_WID_MASK) == 0);
 639         LASSERT ((type & ~IBNAL_WID_MASK) == 0);
 640         return (__u64)(lptr | type);
 641 }
 642
 643 static inline void *
 644 kibnal_wreqid2ptr (__u64 wreqid)
 645 {
 646         return (void *)(((unsigned long)wreqid) & ~IBNAL_WID_MASK);
 647 }
 648
 649 static inline int
 650 kibnal_wreqid2type (__u64 wreqid)
 651 {
 652         return (wreqid & IBNAL_WID_MASK);
 653 }
 654
 655 static inline void
 656 kibnal_set_conn_state (kib_conn_t *conn, int state)
 657 {
 658         CDEBUG(D_NET,"%p state %d\n", conn, state);
 659         conn->ibc_state = state;
 660         mb();
 661 }
 662
 663 #if IBNAL_USE_FMR
 664
 665 static inline int
 666 kibnal_rd_size (kib_rdma_desc_t *rd)
 667 {
 668         return rd->rd_nob;
 669 }
 670
 671 #else
 672 static inline int
 673 kibnal_rd_size (kib_rdma_desc_t *rd)
 674 {
 675         int   i;
 676         int   size;
 677
 678         for (i = size = 0; i < rd->rd_nfrag; i++)
 679                 size += rd->rd_frags[i].rf_nob;
 680
 681         return size;
 682 }
 683 #endif
 684
 685 int  kibnal_startup (lnet_ni_t *ni);
 686 void kibnal_shutdown (lnet_ni_t *ni);
 687 int  kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
 688 int  kibnal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
 689 int  kibnal_eager_recv (lnet_ni_t *ni, void *private,
 690                         lnet_msg_t *lntmsg, void **new_private);
 691 int  kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg,
 692                   int delayed, unsigned int niov,
 693                   struct iovec *iov, lnet_kiov_t *kiov,
 694                   unsigned int offset, unsigned int mlen, unsigned int rlen);
 695 void kibnal_init_msg(kib_msg_t *msg, int type, int body_nob);
 696 void kibnal_pack_msg(kib_msg_t *msg, __u32 version, int credits,
 697                      lnet_nid_t dstnid, __u64 dststamp, __u64 seq);
 698 void kibnal_pack_connmsg(kib_msg_t *msg, __u32 version, int nob, int type,
 699                          lnet_nid_t dstnid, __u64 dststamp);
 700 int  kibnal_unpack_msg(kib_msg_t *msg, __u32 expected_version, int nob);
 701 IB_HANDLE kibnal_create_cep(lnet_nid_t nid);
 702 int  kibnal_create_peer (kib_peer_t **peerp, lnet_nid_t nid);
 703 void kibnal_destroy_peer (kib_peer_t *peer);
 704 kib_peer_t *kibnal_find_peer_locked (lnet_nid_t nid);
 705 int  kibnal_del_peer (lnet_nid_t nid);
 706 void kibnal_peer_alive (kib_peer_t *peer);
 707 void kibnal_unlink_peer_locked (kib_peer_t *peer);
 708 int  kibnal_add_persistent_peer (lnet_nid_t nid);
 709 int  kibnal_close_stale_conns_locked (kib_peer_t *peer,
 710                                       __u64 incarnation);
 711 int  kibnal_conn_rts(kib_conn_t *conn,
 712                      __u32 qpn, __u8 resp_res, __u8 init_depth, __u32 psn);
 713 kib_conn_t *kibnal_create_conn (lnet_nid_t nid, int proto_version);
 714 void kibnal_destroy_conn (kib_conn_t *conn);
 715 void kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg);
 716 int  kibnal_alloc_pages (kib_pages_t **pp, int npages);
 717 void kibnal_free_pages (kib_pages_t *p);
 718 void kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
 719 void kibnal_txlist_done (struct list_head *txlist, int status);
 720 int  kibnal_post_receives (kib_conn_t *conn);
 721 int  kibnal_init_rdma (kib_tx_t *tx, int type, int nob,
 722                        kib_rdma_desc_t *dstrd, __u64 dstcookie);
 723 void kibnal_check_sends (kib_conn_t *conn);
 724 void kibnal_close_conn_locked (kib_conn_t *conn, int error);
 725 int  kibnal_thread_start (int (*fn)(void *arg), void *arg);
 726 int  kibnal_scheduler(void *arg);
 727 int  kibnal_connd (void *arg);
 728 void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob);
 729 void kibnal_close_conn (kib_conn_t *conn, int why);
 730 void kibnal_start_active_rdma (int type, int status,
 731                                kib_rx_t *rx, lnet_msg_t *lntmsg,
 732                                unsigned int niov,
 733                                struct iovec *iov, lnet_kiov_t *kiov,
 734                                unsigned int offset, unsigned int nob);
 735 void kibnal_hca_async_callback (void *hca_arg, IB_EVENT_RECORD *ev);
 736 void kibnal_hca_callback (void *hca_arg, void *cq_arg);
 737 int  kibnal_tunables_init (void);
 738 void kibnal_tunables_fini (void);