release/src-rt-6.x.4708/linux/linux-2.6.36/net/rds/rds.h

   1 #ifndef _RDS_RDS_H
   2 #define _RDS_RDS_H
   3
   4 #include <net/sock.h>
   5 #include <linux/scatterlist.h>
   6 #include <linux/highmem.h>
   7 #include <rdma/rdma_cm.h>
   8 #include <linux/mutex.h>
   9 #include <linux/rds.h>
  10
  11 #include "info.h"
  12
  13 /*
  14  * RDS Network protocol version
  15  */
  16 #define RDS_PROTOCOL_3_0        0x0300
  17 #define RDS_PROTOCOL_3_1        0x0301
  18 #define RDS_PROTOCOL_VERSION    RDS_PROTOCOL_3_1
  19 #define RDS_PROTOCOL_MAJOR(v)   ((v) >> 8)
  20 #define RDS_PROTOCOL_MINOR(v)   ((v) & 255)
  21 #define RDS_PROTOCOL(maj, min)  (((maj) << 8) | min)
  22
  23 #define RDS_PORT        18634
  24
  25 #ifdef ATOMIC64_INIT
  26 #define KERNEL_HAS_ATOMIC64
  27 #endif
  28
  29 #ifdef DEBUG
  30 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
  31 #else
  32 /* sigh, pr_debug() causes unused variable warnings */
  33 static inline void __attribute__ ((format (printf, 1, 2)))
  34 rdsdebug(char *fmt, ...)
  35 {
  36 }
  37 #endif
  38
  39 #define ceil(x, y) \
  40         ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
  41
  42 #define RDS_FRAG_SHIFT  12
  43 #define RDS_FRAG_SIZE   ((unsigned int)(1 << RDS_FRAG_SHIFT))
  44
  45 #define RDS_CONG_MAP_BYTES      (65536 / 8)
  46 #define RDS_CONG_MAP_LONGS      (RDS_CONG_MAP_BYTES / sizeof(unsigned long))
  47 #define RDS_CONG_MAP_PAGES      (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
  48 #define RDS_CONG_MAP_PAGE_BITS  (PAGE_SIZE * 8)
  49
  50 struct rds_cong_map {
  51         struct rb_node          m_rb_node;
  52         __be32                  m_addr;
  53         wait_queue_head_t       m_waitq;
  54         struct list_head        m_conn_list;
  55         unsigned long           m_page_addrs[RDS_CONG_MAP_PAGES];
  56 };
  57
  58
  59 /*
  60  * This is how we will track the connection state:
  61  * A connection is always in one of the following
  62  * states. Updates to the state are atomic and imply
  63  * a memory barrier.
  64  */
  65 enum {
  66         RDS_CONN_DOWN = 0,
  67         RDS_CONN_CONNECTING,
  68         RDS_CONN_DISCONNECTING,
  69         RDS_CONN_UP,
  70         RDS_CONN_ERROR,
  71 };
  72
  73 /* Bits for c_flags */
  74 #define RDS_LL_SEND_FULL        0
  75 #define RDS_RECONNECT_PENDING   1
  76
  77 struct rds_connection {
  78         struct hlist_node       c_hash_node;
  79         __be32                  c_laddr;
  80         __be32                  c_faddr;
  81         unsigned int            c_loopback:1;
  82         struct rds_connection   *c_passive;
  83
  84         struct rds_cong_map     *c_lcong;
  85         struct rds_cong_map     *c_fcong;
  86
  87         struct mutex            c_send_lock;    /* protect send ring */
  88         struct rds_message      *c_xmit_rm;
  89         unsigned long           c_xmit_sg;
  90         unsigned int            c_xmit_hdr_off;
  91         unsigned int            c_xmit_data_off;
  92         unsigned int            c_xmit_rdma_sent;
  93
  94         spinlock_t              c_lock;         /* protect msg queues */
  95         u64                     c_next_tx_seq;
  96         struct list_head        c_send_queue;
  97         struct list_head        c_retrans;
  98
  99         u64                     c_next_rx_seq;
 100
 101         struct rds_transport    *c_trans;
 102         void                    *c_transport_data;
 103
 104         atomic_t                c_state;
 105         unsigned long           c_flags;
 106         unsigned long           c_reconnect_jiffies;
 107         struct delayed_work     c_send_w;
 108         struct delayed_work     c_recv_w;
 109         struct delayed_work     c_conn_w;
 110         struct work_struct      c_down_w;
 111         struct mutex            c_cm_lock;      /* protect conn state & cm */
 112
 113         struct list_head        c_map_item;
 114         unsigned long           c_map_queued;
 115         unsigned long           c_map_offset;
 116         unsigned long           c_map_bytes;
 117
 118         unsigned int            c_unacked_packets;
 119         unsigned int            c_unacked_bytes;
 120
 121         /* Protocol version */
 122         unsigned int            c_version;
 123 };
 124
 125 #define RDS_FLAG_CONG_BITMAP    0x01
 126 #define RDS_FLAG_ACK_REQUIRED   0x02
 127 #define RDS_FLAG_RETRANSMITTED  0x04
 128 #define RDS_MAX_ADV_CREDIT      255
 129
 130 /*
 131  * Maximum space available for extension headers.
 132  */
 133 #define RDS_HEADER_EXT_SPACE    16
 134
 135 struct rds_header {
 136         __be64  h_sequence;
 137         __be64  h_ack;
 138         __be32  h_len;
 139         __be16  h_sport;
 140         __be16  h_dport;
 141         u8      h_flags;
 142         u8      h_credit;
 143         u8      h_padding[4];
 144         __sum16 h_csum;
 145
 146         u8      h_exthdr[RDS_HEADER_EXT_SPACE];
 147 };
 148
 149 /*
 150  * Reserved - indicates end of extensions
 151  */
 152 #define RDS_EXTHDR_NONE         0
 153
 154 /*
 155  * This extension header is included in the very
 156  * first message that is sent on a new connection,
 157  * and identifies the protocol level. This will help
 158  * rolling updates if a future change requires breaking
 159  * the protocol.
 160  * NB: This is no longer true for IB, where we do a version
 161  * negotiation during the connection setup phase (protocol
 162  * version information is included in the RDMA CM private data).
 163  */
 164 #define RDS_EXTHDR_VERSION      1
 165 struct rds_ext_header_version {
 166         __be32                  h_version;
 167 };
 168
 169 /*
 170  * This extension header is included in the RDS message
 171  * chasing an RDMA operation.
 172  */
 173 #define RDS_EXTHDR_RDMA         2
 174 struct rds_ext_header_rdma {
 175         __be32                  h_rdma_rkey;
 176 };
 177
 178 /*
 179  * This extension header tells the peer about the
 180  * destination <R_Key,offset> of the requested RDMA
 181  * operation.
 182  */
 183 #define RDS_EXTHDR_RDMA_DEST    3
 184 struct rds_ext_header_rdma_dest {
 185         __be32                  h_rdma_rkey;
 186         __be32                  h_rdma_offset;
 187 };
 188
 189 #define __RDS_EXTHDR_MAX        16 /* for now */
 190
 191 struct rds_incoming {
 192         atomic_t                i_refcount;
 193         struct list_head        i_item;
 194         struct rds_connection   *i_conn;
 195         struct rds_header       i_hdr;
 196         unsigned long           i_rx_jiffies;
 197         __be32                  i_saddr;
 198
 199         rds_rdma_cookie_t       i_rdma_cookie;
 200 };
 201
 202 /*
 203  * m_sock_item and m_conn_item are on lists that are serialized under
 204  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
 205  * the message will not be put back on the retransmit list after being sent.
 206  * messages that are canceled while being sent rely on this.
 207  *
 208  * m_inc is used by loopback so that it can pass an incoming message straight
 209  * back up into the rx path.  It embeds a wire header which is also used by
 210  * the send path, which is kind of awkward.
 211  *
 212  * m_sock_item indicates the message's presence on a socket's send or receive
 213  * queue.  m_rs will point to that socket.
 214  *
 215  * m_daddr is used by cancellation to prune messages to a given destination.
 216  *
 217  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
 218  * nesting.  As paths iterate over messages on a sock, or conn, they must
 219  * also lock the conn, or sock, to remove the message from those lists too.
 220  * Testing the flag to determine if the message is still on the lists lets
 221  * us avoid testing the list_head directly.  That means each path can use
 222  * the message's list_head to keep it on a local list while juggling locks
 223  * without confusing the other path.
 224  *
 225  * m_ack_seq is an optional field set by transports who need a different
 226  * sequence number range to invalidate.  They can use this in a callback
 227  * that they pass to rds_send_drop_acked() to see if each message has been
 228  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
 229  * had ack_seq set yet.
 230  */
 231 #define RDS_MSG_ON_SOCK         1
 232 #define RDS_MSG_ON_CONN         2
 233 #define RDS_MSG_HAS_ACK_SEQ     3
 234 #define RDS_MSG_ACK_REQUIRED    4
 235 #define RDS_MSG_RETRANSMITTED   5
 236 #define RDS_MSG_MAPPED          6
 237 #define RDS_MSG_PAGEVEC         7
 238
 239 struct rds_message {
 240         atomic_t                m_refcount;
 241         struct list_head        m_sock_item;
 242         struct list_head        m_conn_item;
 243         struct rds_incoming     m_inc;
 244         u64                     m_ack_seq;
 245         __be32                  m_daddr;
 246         unsigned long           m_flags;
 247
 248         /* Never access m_rs without holding m_rs_lock.
 249          * Lock nesting is
 250          *  rm->m_rs_lock
 251          *   -> rs->rs_lock
 252          */
 253         spinlock_t              m_rs_lock;
 254         struct rds_sock         *m_rs;
 255         struct rds_rdma_op      *m_rdma_op;
 256         rds_rdma_cookie_t       m_rdma_cookie;
 257         struct rds_mr           *m_rdma_mr;
 258         unsigned int            m_nents;
 259         unsigned int            m_count;
 260         struct scatterlist      m_sg[0];
 261 };
 262
 263 /*
 264  * The RDS notifier is used (optionally) to tell the application about
 265  * completed RDMA operations. Rather than keeping the whole rds message
 266  * around on the queue, we allocate a small notifier that is put on the
 267  * socket's notifier_list. Notifications are delivered to the application
 268  * through control messages.
 269  */
 270 struct rds_notifier {
 271         struct list_head        n_list;
 272         uint64_t                n_user_token;
 273         int                     n_status;
 274 };
 275
 276
 277 #define RDS_TRANS_IB    0
 278 #define RDS_TRANS_IWARP 1
 279 #define RDS_TRANS_TCP   2
 280 #define RDS_TRANS_COUNT 3
 281
 282 struct rds_transport {
 283         char                    t_name[TRANSNAMSIZ];
 284         struct list_head        t_item;
 285         struct module           *t_owner;
 286         unsigned int            t_prefer_loopback:1;
 287         unsigned int            t_type;
 288
 289         int (*laddr_check)(__be32 addr);
 290         int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
 291         void (*conn_free)(void *data);
 292         int (*conn_connect)(struct rds_connection *conn);
 293         void (*conn_shutdown)(struct rds_connection *conn);
 294         void (*xmit_prepare)(struct rds_connection *conn);
 295         void (*xmit_complete)(struct rds_connection *conn);
 296         int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
 297                     unsigned int hdr_off, unsigned int sg, unsigned int off);
 298         int (*xmit_cong_map)(struct rds_connection *conn,
 299                              struct rds_cong_map *map, unsigned long offset);
 300         int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
 301         int (*recv)(struct rds_connection *conn);
 302         int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
 303                                 size_t size);
 304         void (*inc_purge)(struct rds_incoming *inc);
 305         void (*inc_free)(struct rds_incoming *inc);
 306
 307         int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
 308                                  struct rdma_cm_event *event);
 309         int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
 310         void (*cm_connect_complete)(struct rds_connection *conn,
 311                                     struct rdma_cm_event *event);
 312
 313         unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
 314                                         unsigned int avail);
 315         void (*exit)(void);
 316         void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
 317                         struct rds_sock *rs, u32 *key_ret);
 318         void (*sync_mr)(void *trans_private, int direction);
 319         void (*free_mr)(void *trans_private, int invalidate);
 320         void (*flush_mrs)(void);
 321 };
 322
 323 struct rds_sock {
 324         struct sock             rs_sk;
 325
 326         u64                     rs_user_addr;
 327         u64                     rs_user_bytes;
 328
 329         /*
 330          * bound_addr used for both incoming and outgoing, no INADDR_ANY
 331          * support.
 332          */
 333         struct rb_node          rs_bound_node;
 334         __be32                  rs_bound_addr;
 335         __be32                  rs_conn_addr;
 336         __be16                  rs_bound_port;
 337         __be16                  rs_conn_port;
 338
 339         /*
 340          * This is only used to communicate the transport between bind and
 341          * initiating connections.  All other trans use is referenced through
 342          * the connection.
 343          */
 344         struct rds_transport    *rs_transport;
 345
 346         /*
 347          * rds_sendmsg caches the conn it used the last time around.
 348          * This helps avoid costly lookups.
 349          */
 350         struct rds_connection   *rs_conn;
 351
 352         /* flag indicating we were congested or not */
 353         int                     rs_congested;
 354         /* seen congestion (ENOBUFS) when sending? */
 355         int                     rs_seen_congestion;
 356
 357         /* rs_lock protects all these adjacent members before the newline */
 358         spinlock_t              rs_lock;
 359         struct list_head        rs_send_queue;
 360         u32                     rs_snd_bytes;
 361         int                     rs_rcv_bytes;
 362         struct list_head        rs_notify_queue;        /* currently used for failed RDMAs */
 363
 364         /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
 365          * to decide whether the application should be woken up.
 366          * If not set, we use rs_cong_track to find out whether a cong map
 367          * update arrived.
 368          */
 369         uint64_t                rs_cong_mask;
 370         uint64_t                rs_cong_notify;
 371         struct list_head        rs_cong_list;
 372         unsigned long           rs_cong_track;
 373
 374         /*
 375          * rs_recv_lock protects the receive queue, and is
 376          * used to serialize with rds_release.
 377          */
 378         rwlock_t                rs_recv_lock;
 379         struct list_head        rs_recv_queue;
 380
 381         /* just for stats reporting */
 382         struct list_head        rs_item;
 383
 384         /* these have their own lock */
 385         spinlock_t              rs_rdma_lock;
 386         struct rb_root          rs_rdma_keys;
 387
 388         /* Socket options - in case there will be more */
 389         unsigned char           rs_recverr,
 390                                 rs_cong_monitor;
 391 };
 392
 393 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
 394 {
 395         return container_of(sk, struct rds_sock, rs_sk);
 396 }
 397 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
 398 {
 399         return &rs->rs_sk;
 400 }
 401
 402 /*
 403  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
 404  * to account for overhead.  We don't account for overhead, we just apply
 405  * the number of payload bytes to the specified value.
 406  */
 407 static inline int rds_sk_sndbuf(struct rds_sock *rs)
 408 {
 409         return rds_rs_to_sk(rs)->sk_sndbuf / 2;
 410 }
 411 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
 412 {
 413         return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
 414 }
 415
 416 struct rds_statistics {
 417         uint64_t        s_conn_reset;
 418         uint64_t        s_recv_drop_bad_checksum;
 419         uint64_t        s_recv_drop_old_seq;
 420         uint64_t        s_recv_drop_no_sock;
 421         uint64_t        s_recv_drop_dead_sock;
 422         uint64_t        s_recv_deliver_raced;
 423         uint64_t        s_recv_delivered;
 424         uint64_t        s_recv_queued;
 425         uint64_t        s_recv_immediate_retry;
 426         uint64_t        s_recv_delayed_retry;
 427         uint64_t        s_recv_ack_required;
 428         uint64_t        s_recv_rdma_bytes;
 429         uint64_t        s_recv_ping;
 430         uint64_t        s_send_queue_empty;
 431         uint64_t        s_send_queue_full;
 432         uint64_t        s_send_sem_contention;
 433         uint64_t        s_send_sem_queue_raced;
 434         uint64_t        s_send_immediate_retry;
 435         uint64_t        s_send_delayed_retry;
 436         uint64_t        s_send_drop_acked;
 437         uint64_t        s_send_ack_required;
 438         uint64_t        s_send_queued;
 439         uint64_t        s_send_rdma;
 440         uint64_t        s_send_rdma_bytes;
 441         uint64_t        s_send_pong;
 442         uint64_t        s_page_remainder_hit;
 443         uint64_t        s_page_remainder_miss;
 444         uint64_t        s_copy_to_user;
 445         uint64_t        s_copy_from_user;
 446         uint64_t        s_cong_update_queued;
 447         uint64_t        s_cong_update_received;
 448         uint64_t        s_cong_send_error;
 449         uint64_t        s_cong_send_blocked;
 450 };
 451
 452 /* af_rds.c */
 453 void rds_sock_addref(struct rds_sock *rs);
 454 void rds_sock_put(struct rds_sock *rs);
 455 void rds_wake_sk_sleep(struct rds_sock *rs);
 456 static inline void __rds_wake_sk_sleep(struct sock *sk)
 457 {
 458         wait_queue_head_t *waitq = sk_sleep(sk);
 459
 460         if (!sock_flag(sk, SOCK_DEAD) && waitq)
 461                 wake_up(waitq);
 462 }
 463 extern wait_queue_head_t rds_poll_waitq;
 464
 465
 466 /* bind.c */
 467 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
 468 void rds_remove_bound(struct rds_sock *rs);
 469 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
 470
 471 /* cong.c */
 472 int rds_cong_get_maps(struct rds_connection *conn);
 473 void rds_cong_add_conn(struct rds_connection *conn);
 474 void rds_cong_remove_conn(struct rds_connection *conn);
 475 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
 476 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
 477 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
 478 void rds_cong_queue_updates(struct rds_cong_map *map);
 479 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
 480 int rds_cong_updated_since(unsigned long *recent);
 481 void rds_cong_add_socket(struct rds_sock *);
 482 void rds_cong_remove_socket(struct rds_sock *);
 483 void rds_cong_exit(void);
 484 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
 485
 486 /* conn.c */
 487 int __init rds_conn_init(void);
 488 void rds_conn_exit(void);
 489 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 490                                        struct rds_transport *trans, gfp_t gfp);
 491 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 492                                struct rds_transport *trans, gfp_t gfp);
 493 void rds_conn_destroy(struct rds_connection *conn);
 494 void rds_conn_reset(struct rds_connection *conn);
 495 void rds_conn_drop(struct rds_connection *conn);
 496 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 497                           struct rds_info_iterator *iter,
 498                           struct rds_info_lengths *lens,
 499                           int (*visitor)(struct rds_connection *, void *),
 500                           size_t item_len);
 501 void __rds_conn_error(struct rds_connection *conn, const char *, ...)
 502                                 __attribute__ ((format (printf, 2, 3)));
 503 #define rds_conn_error(conn, fmt...) \
 504         __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
 505
 506 static inline int
 507 rds_conn_transition(struct rds_connection *conn, int old, int new)
 508 {
 509         return atomic_cmpxchg(&conn->c_state, old, new) == old;
 510 }
 511
 512 static inline int
 513 rds_conn_state(struct rds_connection *conn)
 514 {
 515         return atomic_read(&conn->c_state);
 516 }
 517
 518 static inline int
 519 rds_conn_up(struct rds_connection *conn)
 520 {
 521         return atomic_read(&conn->c_state) == RDS_CONN_UP;
 522 }
 523
 524 static inline int
 525 rds_conn_connecting(struct rds_connection *conn)
 526 {
 527         return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
 528 }
 529
 530 /* message.c */
 531 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
 532 struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
 533                                                size_t total_len);
 534 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
 535 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
 536                                  __be16 dport, u64 seq);
 537 int rds_message_add_extension(struct rds_header *hdr,
 538                               unsigned int type, const void *data, unsigned int len);
 539 int rds_message_next_extension(struct rds_header *hdr,
 540                                unsigned int *pos, void *buf, unsigned int *buflen);
 541 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
 542 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
 543 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
 544 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
 545                                  struct iovec *first_iov, size_t size);
 546 void rds_message_inc_purge(struct rds_incoming *inc);
 547 void rds_message_inc_free(struct rds_incoming *inc);
 548 void rds_message_addref(struct rds_message *rm);
 549 void rds_message_put(struct rds_message *rm);
 550 void rds_message_wait(struct rds_message *rm);
 551 void rds_message_unmapped(struct rds_message *rm);
 552
 553 static inline void rds_message_make_checksum(struct rds_header *hdr)
 554 {
 555         hdr->h_csum = 0;
 556         hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
 557 }
 558
 559 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
 560 {
 561         return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
 562 }
 563
 564
 565 /* page.c */
 566 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
 567                              gfp_t gfp);
 568 int rds_page_copy_user(struct page *page, unsigned long offset,
 569                        void __user *ptr, unsigned long bytes,
 570                        int to_user);
 571 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
 572         rds_page_copy_user(page, offset, ptr, bytes, 1)
 573 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
 574         rds_page_copy_user(page, offset, ptr, bytes, 0)
 575 void rds_page_exit(void);
 576
 577 /* recv.c */
 578 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
 579                   __be32 saddr);
 580 void rds_inc_addref(struct rds_incoming *inc);
 581 void rds_inc_put(struct rds_incoming *inc);
 582 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
 583                        struct rds_incoming *inc, gfp_t gfp, enum km_type km);
 584 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 585                 size_t size, int msg_flags);
 586 void rds_clear_recv_queue(struct rds_sock *rs);
 587 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
 588 void rds_inc_info_copy(struct rds_incoming *inc,
 589                        struct rds_info_iterator *iter,
 590                        __be32 saddr, __be32 daddr, int flip);
 591
 592 /* send.c */
 593 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 594                 size_t payload_len);
 595 void rds_send_reset(struct rds_connection *conn);
 596 int rds_send_xmit(struct rds_connection *conn);
 597 struct sockaddr_in;
 598 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
 599 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
 600 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 601                          is_acked_func is_acked);
 602 int rds_send_acked_before(struct rds_connection *conn, u64 seq);
 603 void rds_send_remove_from_sock(struct list_head *messages, int status);
 604 int rds_send_pong(struct rds_connection *conn, __be16 dport);
 605 struct rds_message *rds_send_get_message(struct rds_connection *,
 606                                          struct rds_rdma_op *);
 607
 608 /* rdma.c */
 609 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
 610
 611 /* stats.c */
 612 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
 613 #define rds_stats_inc_which(which, member) do {         \
 614         per_cpu(which, get_cpu()).member++;             \
 615         put_cpu();                                      \
 616 } while (0)
 617 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
 618 #define rds_stats_add_which(which, member, count) do {          \
 619         per_cpu(which, get_cpu()).member += count;      \
 620         put_cpu();                                      \
 621 } while (0)
 622 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
 623 int __init rds_stats_init(void);
 624 void rds_stats_exit(void);
 625 void rds_stats_info_copy(struct rds_info_iterator *iter,
 626                          uint64_t *values, const char *const *names,
 627                          size_t nr);
 628
 629 /* sysctl.c */
 630 int __init rds_sysctl_init(void);
 631 void rds_sysctl_exit(void);
 632 extern unsigned long rds_sysctl_sndbuf_min;
 633 extern unsigned long rds_sysctl_sndbuf_default;
 634 extern unsigned long rds_sysctl_sndbuf_max;
 635 extern unsigned long rds_sysctl_reconnect_min_jiffies;
 636 extern unsigned long rds_sysctl_reconnect_max_jiffies;
 637 extern unsigned int  rds_sysctl_max_unacked_packets;
 638 extern unsigned int  rds_sysctl_max_unacked_bytes;
 639 extern unsigned int  rds_sysctl_ping_enable;
 640 extern unsigned long rds_sysctl_trace_flags;
 641 extern unsigned int  rds_sysctl_trace_level;
 642
 643 /* threads.c */
 644 int __init rds_threads_init(void);
 645 void rds_threads_exit(void);
 646 extern struct workqueue_struct *rds_wq;
 647 void rds_connect_worker(struct work_struct *);
 648 void rds_shutdown_worker(struct work_struct *);
 649 void rds_send_worker(struct work_struct *);
 650 void rds_recv_worker(struct work_struct *);
 651 void rds_connect_complete(struct rds_connection *conn);
 652
 653 /* transport.c */
 654 int rds_trans_register(struct rds_transport *trans);
 655 void rds_trans_unregister(struct rds_transport *trans);
 656 struct rds_transport *rds_trans_get_preferred(__be32 addr);
 657 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
 658                                        unsigned int avail);
 659 int __init rds_trans_init(void);
 660 void rds_trans_exit(void);
 661
 662 #endif