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 /*
  24  * XXX randomly chosen, but at least seems to be unused:
  25  * #               18464-18768 Unassigned
  26  * We should do better.  We want a reserved port to discourage unpriv'ed
  27  * userspace from listening.
  28  */
  29 #define RDS_PORT        18634
  30
  31 #ifdef ATOMIC64_INIT
  32 #define KERNEL_HAS_ATOMIC64
  33 #endif
  34
  35 #ifdef DEBUG
  36 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
  37 #else
  38 /* sigh, pr_debug() causes unused variable warnings */
  39 static inline __printf(1, 2)
  40 void rdsdebug(char *fmt, ...)
  41 {
  42 }
  43 #endif
  44
  45 /* XXX is there one of these somewhere? */
  46 #define ceil(x, y) \
  47         ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
  48
  49 #define RDS_FRAG_SHIFT  12
  50 #define RDS_FRAG_SIZE   ((unsigned int)(1 << RDS_FRAG_SHIFT))
  51
  52 #define RDS_CONG_MAP_BYTES      (65536 / 8)
  53 #define RDS_CONG_MAP_PAGES      (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
  54 #define RDS_CONG_MAP_PAGE_BITS  (PAGE_SIZE * 8)
  55
  56 struct rds_cong_map {
  57         struct rb_node          m_rb_node;
  58         __be32                  m_addr;
  59         wait_queue_head_t       m_waitq;
  60         struct list_head        m_conn_list;
  61         unsigned long           m_page_addrs[RDS_CONG_MAP_PAGES];
  62 };
  63
  64
  65 /*
  66  * This is how we will track the connection state:
  67  * A connection is always in one of the following
  68  * states. Updates to the state are atomic and imply
  69  * a memory barrier.
  70  */
  71 enum {
  72         RDS_CONN_DOWN = 0,
  73         RDS_CONN_CONNECTING,
  74         RDS_CONN_DISCONNECTING,
  75         RDS_CONN_UP,
  76         RDS_CONN_ERROR,
  77 };
  78
  79 /* Bits for c_flags */
  80 #define RDS_LL_SEND_FULL        0
  81 #define RDS_RECONNECT_PENDING   1
  82 #define RDS_IN_XMIT             2
  83
  84 struct rds_connection {
  85         struct hlist_node       c_hash_node;
  86         __be32                  c_laddr;
  87         __be32                  c_faddr;
  88         unsigned int            c_loopback:1;
  89         struct rds_connection   *c_passive;
  90
  91         struct rds_cong_map     *c_lcong;
  92         struct rds_cong_map     *c_fcong;
  93
  94         struct rds_message      *c_xmit_rm;
  95         unsigned long           c_xmit_sg;
  96         unsigned int            c_xmit_hdr_off;
  97         unsigned int            c_xmit_data_off;
  98         unsigned int            c_xmit_atomic_sent;
  99         unsigned int            c_xmit_rdma_sent;
 100         unsigned int            c_xmit_data_sent;
 101
 102         spinlock_t              c_lock;         /* protect msg queues */
 103         u64                     c_next_tx_seq;
 104         struct list_head        c_send_queue;
 105         struct list_head        c_retrans;
 106
 107         u64                     c_next_rx_seq;
 108
 109         struct rds_transport    *c_trans;
 110         void                    *c_transport_data;
 111
 112         atomic_t                c_state;
 113         unsigned long           c_send_gen;
 114         unsigned long           c_flags;
 115         unsigned long           c_reconnect_jiffies;
 116         struct delayed_work     c_send_w;
 117         struct delayed_work     c_recv_w;
 118         struct delayed_work     c_conn_w;
 119         struct work_struct      c_down_w;
 120         struct mutex            c_cm_lock;      /* protect conn state & cm */
 121         wait_queue_head_t       c_waitq;
 122
 123         struct list_head        c_map_item;
 124         unsigned long           c_map_queued;
 125
 126         unsigned int            c_unacked_packets;
 127         unsigned int            c_unacked_bytes;
 128
 129         /* Protocol version */
 130         unsigned int            c_version;
 131         possible_net_t          c_net;
 132 };
 133
 134 static inline
 135 struct net *rds_conn_net(struct rds_connection *conn)
 136 {
 137         return read_pnet(&conn->c_net);
 138 }
 139
 140 static inline
 141 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
 142 {
 143         write_pnet(&conn->c_net, net);
 144 }
 145
 146 #define RDS_FLAG_CONG_BITMAP    0x01
 147 #define RDS_FLAG_ACK_REQUIRED   0x02
 148 #define RDS_FLAG_RETRANSMITTED  0x04
 149 #define RDS_MAX_ADV_CREDIT      255
 150
 151 /*
 152  * Maximum space available for extension headers.
 153  */
 154 #define RDS_HEADER_EXT_SPACE    16
 155
 156 struct rds_header {
 157         __be64  h_sequence;
 158         __be64  h_ack;
 159         __be32  h_len;
 160         __be16  h_sport;
 161         __be16  h_dport;
 162         u8      h_flags;
 163         u8      h_credit;
 164         u8      h_padding[4];
 165         __sum16 h_csum;
 166
 167         u8      h_exthdr[RDS_HEADER_EXT_SPACE];
 168 };
 169
 170 /*
 171  * Reserved - indicates end of extensions
 172  */
 173 #define RDS_EXTHDR_NONE         0
 174
 175 /*
 176  * This extension header is included in the very
 177  * first message that is sent on a new connection,
 178  * and identifies the protocol level. This will help
 179  * rolling updates if a future change requires breaking
 180  * the protocol.
 181  * NB: This is no longer true for IB, where we do a version
 182  * negotiation during the connection setup phase (protocol
 183  * version information is included in the RDMA CM private data).
 184  */
 185 #define RDS_EXTHDR_VERSION      1
 186 struct rds_ext_header_version {
 187         __be32                  h_version;
 188 };
 189
 190 /*
 191  * This extension header is included in the RDS message
 192  * chasing an RDMA operation.
 193  */
 194 #define RDS_EXTHDR_RDMA         2
 195 struct rds_ext_header_rdma {
 196         __be32                  h_rdma_rkey;
 197 };
 198
 199 /*
 200  * This extension header tells the peer about the
 201  * destination <R_Key,offset> of the requested RDMA
 202  * operation.
 203  */
 204 #define RDS_EXTHDR_RDMA_DEST    3
 205 struct rds_ext_header_rdma_dest {
 206         __be32                  h_rdma_rkey;
 207         __be32                  h_rdma_offset;
 208 };
 209
 210 #define __RDS_EXTHDR_MAX        16 /* for now */
 211
 212 struct rds_incoming {
 213         atomic_t                i_refcount;
 214         struct list_head        i_item;
 215         struct rds_connection   *i_conn;
 216         struct rds_header       i_hdr;
 217         unsigned long           i_rx_jiffies;
 218         __be32                  i_saddr;
 219
 220         rds_rdma_cookie_t       i_rdma_cookie;
 221 };
 222
 223 struct rds_mr {
 224         struct rb_node          r_rb_node;
 225         atomic_t                r_refcount;
 226         u32                     r_key;
 227
 228         /* A copy of the creation flags */
 229         unsigned int            r_use_once:1;
 230         unsigned int            r_invalidate:1;
 231         unsigned int            r_write:1;
 232
 233         /* This is for RDS_MR_DEAD.
 234          * It would be nice & consistent to make this part of the above
 235          * bit field here, but we need to use test_and_set_bit.
 236          */
 237         unsigned long           r_state;
 238         struct rds_sock         *r_sock; /* back pointer to the socket that owns us */
 239         struct rds_transport    *r_trans;
 240         void                    *r_trans_private;
 241 };
 242
 243 /* Flags for mr->r_state */
 244 #define RDS_MR_DEAD             0
 245
 246 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
 247 {
 248         return r_key | (((u64) offset) << 32);
 249 }
 250
 251 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
 252 {
 253         return cookie;
 254 }
 255
 256 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
 257 {
 258         return cookie >> 32;
 259 }
 260
 261 /* atomic operation types */
 262 #define RDS_ATOMIC_TYPE_CSWP            0
 263 #define RDS_ATOMIC_TYPE_FADD            1
 264
 265 /*
 266  * m_sock_item and m_conn_item are on lists that are serialized under
 267  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
 268  * the message will not be put back on the retransmit list after being sent.
 269  * messages that are canceled while being sent rely on this.
 270  *
 271  * m_inc is used by loopback so that it can pass an incoming message straight
 272  * back up into the rx path.  It embeds a wire header which is also used by
 273  * the send path, which is kind of awkward.
 274  *
 275  * m_sock_item indicates the message's presence on a socket's send or receive
 276  * queue.  m_rs will point to that socket.
 277  *
 278  * m_daddr is used by cancellation to prune messages to a given destination.
 279  *
 280  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
 281  * nesting.  As paths iterate over messages on a sock, or conn, they must
 282  * also lock the conn, or sock, to remove the message from those lists too.
 283  * Testing the flag to determine if the message is still on the lists lets
 284  * us avoid testing the list_head directly.  That means each path can use
 285  * the message's list_head to keep it on a local list while juggling locks
 286  * without confusing the other path.
 287  *
 288  * m_ack_seq is an optional field set by transports who need a different
 289  * sequence number range to invalidate.  They can use this in a callback
 290  * that they pass to rds_send_drop_acked() to see if each message has been
 291  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
 292  * had ack_seq set yet.
 293  */
 294 #define RDS_MSG_ON_SOCK         1
 295 #define RDS_MSG_ON_CONN         2
 296 #define RDS_MSG_HAS_ACK_SEQ     3
 297 #define RDS_MSG_ACK_REQUIRED    4
 298 #define RDS_MSG_RETRANSMITTED   5
 299 #define RDS_MSG_MAPPED          6
 300 #define RDS_MSG_PAGEVEC         7
 301
 302 struct rds_message {
 303         atomic_t                m_refcount;
 304         struct list_head        m_sock_item;
 305         struct list_head        m_conn_item;
 306         struct rds_incoming     m_inc;
 307         u64                     m_ack_seq;
 308         __be32                  m_daddr;
 309         unsigned long           m_flags;
 310
 311         /* Never access m_rs without holding m_rs_lock.
 312          * Lock nesting is
 313          *  rm->m_rs_lock
 314          *   -> rs->rs_lock
 315          */
 316         spinlock_t              m_rs_lock;
 317         wait_queue_head_t       m_flush_wait;
 318
 319         struct rds_sock         *m_rs;
 320
 321         /* cookie to send to remote, in rds header */
 322         rds_rdma_cookie_t       m_rdma_cookie;
 323
 324         unsigned int            m_used_sgs;
 325         unsigned int            m_total_sgs;
 326
 327         void                    *m_final_op;
 328
 329         struct {
 330                 struct rm_atomic_op {
 331                         int                     op_type;
 332                         union {
 333                                 struct {
 334                                         uint64_t        compare;
 335                                         uint64_t        swap;
 336                                         uint64_t        compare_mask;
 337                                         uint64_t        swap_mask;
 338                                 } op_m_cswp;
 339                                 struct {
 340                                         uint64_t        add;
 341                                         uint64_t        nocarry_mask;
 342                                 } op_m_fadd;
 343                         };
 344
 345                         u32                     op_rkey;
 346                         u64                     op_remote_addr;
 347                         unsigned int            op_notify:1;
 348                         unsigned int            op_recverr:1;
 349                         unsigned int            op_mapped:1;
 350                         unsigned int            op_silent:1;
 351                         unsigned int            op_active:1;
 352                         struct scatterlist      *op_sg;
 353                         struct rds_notifier     *op_notifier;
 354
 355                         struct rds_mr           *op_rdma_mr;
 356                 } atomic;
 357                 struct rm_rdma_op {
 358                         u32                     op_rkey;
 359                         u64                     op_remote_addr;
 360                         unsigned int            op_write:1;
 361                         unsigned int            op_fence:1;
 362                         unsigned int            op_notify:1;
 363                         unsigned int            op_recverr:1;
 364                         unsigned int            op_mapped:1;
 365                         unsigned int            op_silent:1;
 366                         unsigned int            op_active:1;
 367                         unsigned int            op_bytes;
 368                         unsigned int            op_nents;
 369                         unsigned int            op_count;
 370                         struct scatterlist      *op_sg;
 371                         struct rds_notifier     *op_notifier;
 372
 373                         struct rds_mr           *op_rdma_mr;
 374                 } rdma;
 375                 struct rm_data_op {
 376                         unsigned int            op_active:1;
 377                         unsigned int            op_nents;
 378                         unsigned int            op_count;
 379                         unsigned int            op_dmasg;
 380                         unsigned int            op_dmaoff;
 381                         struct scatterlist      *op_sg;
 382                 } data;
 383         };
 384 };
 385
 386 /*
 387  * The RDS notifier is used (optionally) to tell the application about
 388  * completed RDMA operations. Rather than keeping the whole rds message
 389  * around on the queue, we allocate a small notifier that is put on the
 390  * socket's notifier_list. Notifications are delivered to the application
 391  * through control messages.
 392  */
 393 struct rds_notifier {
 394         struct list_head        n_list;
 395         uint64_t                n_user_token;
 396         int                     n_status;
 397 };
 398
 399 /**
 400  * struct rds_transport -  transport specific behavioural hooks
 401  *
 402  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
 403  *        part of a message.  The caller serializes on the send_sem so this
 404  *        doesn't need to be reentrant for a given conn.  The header must be
 405  *        sent before the data payload.  .xmit must be prepared to send a
 406  *        message with no data payload.  .xmit should return the number of
 407  *        bytes that were sent down the connection, including header bytes.
 408  *        Returning 0 tells the caller that it doesn't need to perform any
 409  *        additional work now.  This is usually the case when the transport has
 410  *        filled the sending queue for its connection and will handle
 411  *        triggering the rds thread to continue the send when space becomes
 412  *        available.  Returning -EAGAIN tells the caller to retry the send
 413  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
 414  *        some point in the future.
 415  *
 416  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
 417  *                 it returns the connection can not call rds_recv_incoming().
 418  *                 This will only be called once after conn_connect returns
 419  *                 non-zero success and will The caller serializes this with
 420  *                 the send and connecting paths (xmit_* and conn_*).  The
 421  *                 transport is responsible for other serialization, including
 422  *                 rds_recv_incoming().  This is called in process context but
 423  *                 should try hard not to block.
 424  */
 425
 426 struct rds_transport {
 427         char                    t_name[TRANSNAMSIZ];
 428         struct list_head        t_item;
 429         struct module           *t_owner;
 430         unsigned int            t_prefer_loopback:1;
 431         unsigned int            t_type;
 432
 433         int (*laddr_check)(struct net *net, __be32 addr);
 434         int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
 435         void (*conn_free)(void *data);
 436         int (*conn_connect)(struct rds_connection *conn);
 437         void (*conn_shutdown)(struct rds_connection *conn);
 438         void (*xmit_prepare)(struct rds_connection *conn);
 439         void (*xmit_complete)(struct rds_connection *conn);
 440         int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
 441                     unsigned int hdr_off, unsigned int sg, unsigned int off);
 442         int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
 443         int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
 444         int (*recv)(struct rds_connection *conn);
 445         int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
 446         void (*inc_free)(struct rds_incoming *inc);
 447
 448         int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
 449                                  struct rdma_cm_event *event);
 450         int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
 451         void (*cm_connect_complete)(struct rds_connection *conn,
 452                                     struct rdma_cm_event *event);
 453
 454         unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
 455                                         unsigned int avail);
 456         void (*exit)(void);
 457         void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
 458                         struct rds_sock *rs, u32 *key_ret);
 459         void (*sync_mr)(void *trans_private, int direction);
 460         void (*free_mr)(void *trans_private, int invalidate);
 461         void (*flush_mrs)(void);
 462 };
 463
 464 struct rds_sock {
 465         struct sock             rs_sk;
 466
 467         u64                     rs_user_addr;
 468         u64                     rs_user_bytes;
 469
 470         /*
 471          * bound_addr used for both incoming and outgoing, no INADDR_ANY
 472          * support.
 473          */
 474         struct hlist_node       rs_bound_node;
 475         __be32                  rs_bound_addr;
 476         __be32                  rs_conn_addr;
 477         __be16                  rs_bound_port;
 478         __be16                  rs_conn_port;
 479         struct rds_transport    *rs_transport;
 480
 481         /*
 482          * rds_sendmsg caches the conn it used the last time around.
 483          * This helps avoid costly lookups.
 484          */
 485         struct rds_connection   *rs_conn;
 486
 487         /* flag indicating we were congested or not */
 488         int                     rs_congested;
 489         /* seen congestion (ENOBUFS) when sending? */
 490         int                     rs_seen_congestion;
 491
 492         /* rs_lock protects all these adjacent members before the newline */
 493         spinlock_t              rs_lock;
 494         struct list_head        rs_send_queue;
 495         u32                     rs_snd_bytes;
 496         int                     rs_rcv_bytes;
 497         struct list_head        rs_notify_queue;        /* currently used for failed RDMAs */
 498
 499         /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
 500          * to decide whether the application should be woken up.
 501          * If not set, we use rs_cong_track to find out whether a cong map
 502          * update arrived.
 503          */
 504         uint64_t                rs_cong_mask;
 505         uint64_t                rs_cong_notify;
 506         struct list_head        rs_cong_list;
 507         unsigned long           rs_cong_track;
 508
 509         /*
 510          * rs_recv_lock protects the receive queue, and is
 511          * used to serialize with rds_release.
 512          */
 513         rwlock_t                rs_recv_lock;
 514         struct list_head        rs_recv_queue;
 515
 516         /* just for stats reporting */
 517         struct list_head        rs_item;
 518
 519         /* these have their own lock */
 520         spinlock_t              rs_rdma_lock;
 521         struct rb_root          rs_rdma_keys;
 522
 523         /* Socket options - in case there will be more */
 524         unsigned char           rs_recverr,
 525                                 rs_cong_monitor;
 526 };
 527
 528 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
 529 {
 530         return container_of(sk, struct rds_sock, rs_sk);
 531 }
 532 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
 533 {
 534         return &rs->rs_sk;
 535 }
 536
 537 /*
 538  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
 539  * to account for overhead.  We don't account for overhead, we just apply
 540  * the number of payload bytes to the specified value.
 541  */
 542 static inline int rds_sk_sndbuf(struct rds_sock *rs)
 543 {
 544         return rds_rs_to_sk(rs)->sk_sndbuf / 2;
 545 }
 546 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
 547 {
 548         return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
 549 }
 550
 551 struct rds_statistics {
 552         uint64_t        s_conn_reset;
 553         uint64_t        s_recv_drop_bad_checksum;
 554         uint64_t        s_recv_drop_old_seq;
 555         uint64_t        s_recv_drop_no_sock;
 556         uint64_t        s_recv_drop_dead_sock;
 557         uint64_t        s_recv_deliver_raced;
 558         uint64_t        s_recv_delivered;
 559         uint64_t        s_recv_queued;
 560         uint64_t        s_recv_immediate_retry;
 561         uint64_t        s_recv_delayed_retry;
 562         uint64_t        s_recv_ack_required;
 563         uint64_t        s_recv_rdma_bytes;
 564         uint64_t        s_recv_ping;
 565         uint64_t        s_send_queue_empty;
 566         uint64_t        s_send_queue_full;
 567         uint64_t        s_send_lock_contention;
 568         uint64_t        s_send_lock_queue_raced;
 569         uint64_t        s_send_immediate_retry;
 570         uint64_t        s_send_delayed_retry;
 571         uint64_t        s_send_drop_acked;
 572         uint64_t        s_send_ack_required;
 573         uint64_t        s_send_queued;
 574         uint64_t        s_send_rdma;
 575         uint64_t        s_send_rdma_bytes;
 576         uint64_t        s_send_pong;
 577         uint64_t        s_page_remainder_hit;
 578         uint64_t        s_page_remainder_miss;
 579         uint64_t        s_copy_to_user;
 580         uint64_t        s_copy_from_user;
 581         uint64_t        s_cong_update_queued;
 582         uint64_t        s_cong_update_received;
 583         uint64_t        s_cong_send_error;
 584         uint64_t        s_cong_send_blocked;
 585 };
 586
 587 /* af_rds.c */
 588 void rds_sock_addref(struct rds_sock *rs);
 589 void rds_sock_put(struct rds_sock *rs);
 590 void rds_wake_sk_sleep(struct rds_sock *rs);
 591 static inline void __rds_wake_sk_sleep(struct sock *sk)
 592 {
 593         wait_queue_head_t *waitq = sk_sleep(sk);
 594
 595         if (!sock_flag(sk, SOCK_DEAD) && waitq)
 596                 wake_up(waitq);
 597 }
 598 extern wait_queue_head_t rds_poll_waitq;
 599
 600
 601 /* bind.c */
 602 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
 603 void rds_remove_bound(struct rds_sock *rs);
 604 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
 605
 606 /* cong.c */
 607 int rds_cong_get_maps(struct rds_connection *conn);
 608 void rds_cong_add_conn(struct rds_connection *conn);
 609 void rds_cong_remove_conn(struct rds_connection *conn);
 610 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
 611 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
 612 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
 613 void rds_cong_queue_updates(struct rds_cong_map *map);
 614 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
 615 int rds_cong_updated_since(unsigned long *recent);
 616 void rds_cong_add_socket(struct rds_sock *);
 617 void rds_cong_remove_socket(struct rds_sock *);
 618 void rds_cong_exit(void);
 619 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
 620
 621 /* conn.c */
 622 int rds_conn_init(void);
 623 void rds_conn_exit(void);
 624 struct rds_connection *rds_conn_create(struct net *net,
 625                                        __be32 laddr, __be32 faddr,
 626                                        struct rds_transport *trans, gfp_t gfp);
 627 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 628                                                 __be32 laddr, __be32 faddr,
 629                                struct rds_transport *trans, gfp_t gfp);
 630 void rds_conn_shutdown(struct rds_connection *conn);
 631 void rds_conn_destroy(struct rds_connection *conn);
 632 void rds_conn_drop(struct rds_connection *conn);
 633 void rds_conn_connect_if_down(struct rds_connection *conn);
 634 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 635                           struct rds_info_iterator *iter,
 636                           struct rds_info_lengths *lens,
 637                           int (*visitor)(struct rds_connection *, void *),
 638                           size_t item_len);
 639 __printf(2, 3)
 640 void __rds_conn_error(struct rds_connection *conn, const char *, ...);
 641 #define rds_conn_error(conn, fmt...) \
 642         __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
 643
 644 static inline int
 645 rds_conn_transition(struct rds_connection *conn, int old, int new)
 646 {
 647         return atomic_cmpxchg(&conn->c_state, old, new) == old;
 648 }
 649
 650 static inline int
 651 rds_conn_state(struct rds_connection *conn)
 652 {
 653         return atomic_read(&conn->c_state);
 654 }
 655
 656 static inline int
 657 rds_conn_up(struct rds_connection *conn)
 658 {
 659         return atomic_read(&conn->c_state) == RDS_CONN_UP;
 660 }
 661
 662 static inline int
 663 rds_conn_connecting(struct rds_connection *conn)
 664 {
 665         return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
 666 }
 667
 668 /* message.c */
 669 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
 670 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
 671 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
 672 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
 673 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
 674                                  __be16 dport, u64 seq);
 675 int rds_message_add_extension(struct rds_header *hdr,
 676                               unsigned int type, const void *data, unsigned int len);
 677 int rds_message_next_extension(struct rds_header *hdr,
 678                                unsigned int *pos, void *buf, unsigned int *buflen);
 679 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
 680 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
 681 void rds_message_inc_free(struct rds_incoming *inc);
 682 void rds_message_addref(struct rds_message *rm);
 683 void rds_message_put(struct rds_message *rm);
 684 void rds_message_wait(struct rds_message *rm);
 685 void rds_message_unmapped(struct rds_message *rm);
 686
 687 static inline void rds_message_make_checksum(struct rds_header *hdr)
 688 {
 689         hdr->h_csum = 0;
 690         hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
 691 }
 692
 693 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
 694 {
 695         return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
 696 }
 697
 698
 699 /* page.c */
 700 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
 701                              gfp_t gfp);
 702 int rds_page_copy_user(struct page *page, unsigned long offset,
 703                        void __user *ptr, unsigned long bytes,
 704                        int to_user);
 705 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
 706         rds_page_copy_user(page, offset, ptr, bytes, 1)
 707 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
 708         rds_page_copy_user(page, offset, ptr, bytes, 0)
 709 void rds_page_exit(void);
 710
 711 /* recv.c */
 712 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
 713                   __be32 saddr);
 714 void rds_inc_put(struct rds_incoming *inc);
 715 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
 716                        struct rds_incoming *inc, gfp_t gfp);
 717 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
 718                 int msg_flags);
 719 void rds_clear_recv_queue(struct rds_sock *rs);
 720 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
 721 void rds_inc_info_copy(struct rds_incoming *inc,
 722                        struct rds_info_iterator *iter,
 723                        __be32 saddr, __be32 daddr, int flip);
 724
 725 /* send.c */
 726 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
 727 void rds_send_reset(struct rds_connection *conn);
 728 int rds_send_xmit(struct rds_connection *conn);
 729 struct sockaddr_in;
 730 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
 731 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
 732 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 733                          is_acked_func is_acked);
 734 int rds_send_pong(struct rds_connection *conn, __be16 dport);
 735 struct rds_message *rds_send_get_message(struct rds_connection *,
 736                                          struct rm_rdma_op *);
 737
 738 /* rdma.c */
 739 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
 740 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
 741 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
 742 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
 743 void rds_rdma_drop_keys(struct rds_sock *rs);
 744 int rds_rdma_extra_size(struct rds_rdma_args *args);
 745 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 746                           struct cmsghdr *cmsg);
 747 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
 748                           struct cmsghdr *cmsg);
 749 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 750                           struct cmsghdr *cmsg);
 751 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
 752                           struct cmsghdr *cmsg);
 753 void rds_rdma_free_op(struct rm_rdma_op *ro);
 754 void rds_atomic_free_op(struct rm_atomic_op *ao);
 755 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
 756 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
 757 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
 758                     struct cmsghdr *cmsg);
 759
 760 void __rds_put_mr_final(struct rds_mr *mr);
 761 static inline void rds_mr_put(struct rds_mr *mr)
 762 {
 763         if (atomic_dec_and_test(&mr->r_refcount))
 764                 __rds_put_mr_final(mr);
 765 }
 766
 767 /* stats.c */
 768 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
 769 #define rds_stats_inc_which(which, member) do {         \
 770         per_cpu(which, get_cpu()).member++;             \
 771         put_cpu();                                      \
 772 } while (0)
 773 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
 774 #define rds_stats_add_which(which, member, count) do {          \
 775         per_cpu(which, get_cpu()).member += count;      \
 776         put_cpu();                                      \
 777 } while (0)
 778 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
 779 int rds_stats_init(void);
 780 void rds_stats_exit(void);
 781 void rds_stats_info_copy(struct rds_info_iterator *iter,
 782                          uint64_t *values, const char *const *names,
 783                          size_t nr);
 784
 785 /* sysctl.c */
 786 int rds_sysctl_init(void);
 787 void rds_sysctl_exit(void);
 788 extern unsigned long rds_sysctl_sndbuf_min;
 789 extern unsigned long rds_sysctl_sndbuf_default;
 790 extern unsigned long rds_sysctl_sndbuf_max;
 791 extern unsigned long rds_sysctl_reconnect_min_jiffies;
 792 extern unsigned long rds_sysctl_reconnect_max_jiffies;
 793 extern unsigned int  rds_sysctl_max_unacked_packets;
 794 extern unsigned int  rds_sysctl_max_unacked_bytes;
 795 extern unsigned int  rds_sysctl_ping_enable;
 796 extern unsigned long rds_sysctl_trace_flags;
 797 extern unsigned int  rds_sysctl_trace_level;
 798
 799 /* threads.c */
 800 int rds_threads_init(void);
 801 void rds_threads_exit(void);
 802 extern struct workqueue_struct *rds_wq;
 803 void rds_queue_reconnect(struct rds_connection *conn);
 804 void rds_connect_worker(struct work_struct *);
 805 void rds_shutdown_worker(struct work_struct *);
 806 void rds_send_worker(struct work_struct *);
 807 void rds_recv_worker(struct work_struct *);
 808 void rds_connect_complete(struct rds_connection *conn);
 809
 810 /* transport.c */
 811 int rds_trans_register(struct rds_transport *trans);
 812 void rds_trans_unregister(struct rds_transport *trans);
 813 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
 814 void rds_trans_put(struct rds_transport *trans);
 815 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
 816                                        unsigned int avail);
 817 struct rds_transport *rds_trans_get(int t_type);
 818 int rds_trans_init(void);
 819 void rds_trans_exit(void);
 820
 821 #endif