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 #define RDS_RECV_REFILL 3 84 85 struct rds_connection { 86 struct hlist_node c_hash_node; 87 __be32 c_laddr; 88 __be32 c_faddr; 89 unsigned int c_loopback:1, 90 c_outgoing:1, 91 c_pad_to_32:30; 92 struct rds_connection *c_passive; 93 94 struct rds_cong_map *c_lcong; 95 struct rds_cong_map *c_fcong; 96 97 struct rds_message *c_xmit_rm; 98 unsigned long c_xmit_sg; 99 unsigned int c_xmit_hdr_off; 100 unsigned int c_xmit_data_off; 101 unsigned int c_xmit_atomic_sent; 102 unsigned int c_xmit_rdma_sent; 103 unsigned int c_xmit_data_sent; 104 105 spinlock_t c_lock; /* protect msg queues */ 106 u64 c_next_tx_seq; 107 struct list_head c_send_queue; 108 struct list_head c_retrans; 109 110 u64 c_next_rx_seq; 111 112 struct rds_transport *c_trans; 113 void *c_transport_data; 114 115 atomic_t c_state; 116 unsigned long c_send_gen; 117 unsigned long c_flags; 118 unsigned long c_reconnect_jiffies; 119 struct delayed_work c_send_w; 120 struct delayed_work c_recv_w; 121 struct delayed_work c_conn_w; 122 struct work_struct c_down_w; 123 struct mutex c_cm_lock; /* protect conn state & cm */ 124 wait_queue_head_t c_waitq; 125 126 struct list_head c_map_item; 127 unsigned long c_map_queued; 128 129 unsigned int c_unacked_packets; 130 unsigned int c_unacked_bytes; 131 132 /* Protocol version */ 133 unsigned int c_version; 134 possible_net_t c_net; 135 }; 136 137 static inline 138 struct net *rds_conn_net(struct rds_connection *conn) 139 { 140 return read_pnet(&conn->c_net); 141 } 142 143 static inline 144 void rds_conn_net_set(struct rds_connection *conn, struct net *net) 145 { 146 write_pnet(&conn->c_net, net); 147 } 148 149 #define RDS_FLAG_CONG_BITMAP 0x01 150 #define RDS_FLAG_ACK_REQUIRED 0x02 151 #define RDS_FLAG_RETRANSMITTED 0x04 152 #define RDS_MAX_ADV_CREDIT 255 153 154 /* 155 * Maximum space available for extension headers. 156 */ 157 #define RDS_HEADER_EXT_SPACE 16 158 159 struct rds_header { 160 __be64 h_sequence; 161 __be64 h_ack; 162 __be32 h_len; 163 __be16 h_sport; 164 __be16 h_dport; 165 u8 h_flags; 166 u8 h_credit; 167 u8 h_padding[4]; 168 __sum16 h_csum; 169 170 u8 h_exthdr[RDS_HEADER_EXT_SPACE]; 171 }; 172 173 /* 174 * Reserved - indicates end of extensions 175 */ 176 #define RDS_EXTHDR_NONE 0 177 178 /* 179 * This extension header is included in the very 180 * first message that is sent on a new connection, 181 * and identifies the protocol level. This will help 182 * rolling updates if a future change requires breaking 183 * the protocol. 184 * NB: This is no longer true for IB, where we do a version 185 * negotiation during the connection setup phase (protocol 186 * version information is included in the RDMA CM private data). 187 */ 188 #define RDS_EXTHDR_VERSION 1 189 struct rds_ext_header_version { 190 __be32 h_version; 191 }; 192 193 /* 194 * This extension header is included in the RDS message 195 * chasing an RDMA operation. 196 */ 197 #define RDS_EXTHDR_RDMA 2 198 struct rds_ext_header_rdma { 199 __be32 h_rdma_rkey; 200 }; 201 202 /* 203 * This extension header tells the peer about the 204 * destination <R_Key,offset> of the requested RDMA 205 * operation. 206 */ 207 #define RDS_EXTHDR_RDMA_DEST 3 208 struct rds_ext_header_rdma_dest { 209 __be32 h_rdma_rkey; 210 __be32 h_rdma_offset; 211 }; 212 213 #define __RDS_EXTHDR_MAX 16 /* for now */ 214 215 struct rds_incoming { 216 atomic_t i_refcount; 217 struct list_head i_item; 218 struct rds_connection *i_conn; 219 struct rds_header i_hdr; 220 unsigned long i_rx_jiffies; 221 __be32 i_saddr; 222 223 rds_rdma_cookie_t i_rdma_cookie; 224 }; 225 226 struct rds_mr { 227 struct rb_node r_rb_node; 228 atomic_t r_refcount; 229 u32 r_key; 230 231 /* A copy of the creation flags */ 232 unsigned int r_use_once:1; 233 unsigned int r_invalidate:1; 234 unsigned int r_write:1; 235 236 /* This is for RDS_MR_DEAD. 237 * It would be nice & consistent to make this part of the above 238 * bit field here, but we need to use test_and_set_bit. 239 */ 240 unsigned long r_state; 241 struct rds_sock *r_sock; /* back pointer to the socket that owns us */ 242 struct rds_transport *r_trans; 243 void *r_trans_private; 244 }; 245 246 /* Flags for mr->r_state */ 247 #define RDS_MR_DEAD 0 248 249 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) 250 { 251 return r_key | (((u64) offset) << 32); 252 } 253 254 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) 255 { 256 return cookie; 257 } 258 259 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) 260 { 261 return cookie >> 32; 262 } 263 264 /* atomic operation types */ 265 #define RDS_ATOMIC_TYPE_CSWP 0 266 #define RDS_ATOMIC_TYPE_FADD 1 267 268 /* 269 * m_sock_item and m_conn_item are on lists that are serialized under 270 * conn->c_lock. m_sock_item has additional meaning in that once it is empty 271 * the message will not be put back on the retransmit list after being sent. 272 * messages that are canceled while being sent rely on this. 273 * 274 * m_inc is used by loopback so that it can pass an incoming message straight 275 * back up into the rx path. It embeds a wire header which is also used by 276 * the send path, which is kind of awkward. 277 * 278 * m_sock_item indicates the message's presence on a socket's send or receive 279 * queue. m_rs will point to that socket. 280 * 281 * m_daddr is used by cancellation to prune messages to a given destination. 282 * 283 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock 284 * nesting. As paths iterate over messages on a sock, or conn, they must 285 * also lock the conn, or sock, to remove the message from those lists too. 286 * Testing the flag to determine if the message is still on the lists lets 287 * us avoid testing the list_head directly. That means each path can use 288 * the message's list_head to keep it on a local list while juggling locks 289 * without confusing the other path. 290 * 291 * m_ack_seq is an optional field set by transports who need a different 292 * sequence number range to invalidate. They can use this in a callback 293 * that they pass to rds_send_drop_acked() to see if each message has been 294 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't 295 * had ack_seq set yet. 296 */ 297 #define RDS_MSG_ON_SOCK 1 298 #define RDS_MSG_ON_CONN 2 299 #define RDS_MSG_HAS_ACK_SEQ 3 300 #define RDS_MSG_ACK_REQUIRED 4 301 #define RDS_MSG_RETRANSMITTED 5 302 #define RDS_MSG_MAPPED 6 303 #define RDS_MSG_PAGEVEC 7 304 305 struct rds_message { 306 atomic_t m_refcount; 307 struct list_head m_sock_item; 308 struct list_head m_conn_item; 309 struct rds_incoming m_inc; 310 u64 m_ack_seq; 311 __be32 m_daddr; 312 unsigned long m_flags; 313 314 /* Never access m_rs without holding m_rs_lock. 315 * Lock nesting is 316 * rm->m_rs_lock 317 * -> rs->rs_lock 318 */ 319 spinlock_t m_rs_lock; 320 wait_queue_head_t m_flush_wait; 321 322 struct rds_sock *m_rs; 323 324 /* cookie to send to remote, in rds header */ 325 rds_rdma_cookie_t m_rdma_cookie; 326 327 unsigned int m_used_sgs; 328 unsigned int m_total_sgs; 329 330 void *m_final_op; 331 332 struct { 333 struct rm_atomic_op { 334 int op_type; 335 union { 336 struct { 337 uint64_t compare; 338 uint64_t swap; 339 uint64_t compare_mask; 340 uint64_t swap_mask; 341 } op_m_cswp; 342 struct { 343 uint64_t add; 344 uint64_t nocarry_mask; 345 } op_m_fadd; 346 }; 347 348 u32 op_rkey; 349 u64 op_remote_addr; 350 unsigned int op_notify:1; 351 unsigned int op_recverr:1; 352 unsigned int op_mapped:1; 353 unsigned int op_silent:1; 354 unsigned int op_active:1; 355 struct scatterlist *op_sg; 356 struct rds_notifier *op_notifier; 357 358 struct rds_mr *op_rdma_mr; 359 } atomic; 360 struct rm_rdma_op { 361 u32 op_rkey; 362 u64 op_remote_addr; 363 unsigned int op_write:1; 364 unsigned int op_fence:1; 365 unsigned int op_notify:1; 366 unsigned int op_recverr:1; 367 unsigned int op_mapped:1; 368 unsigned int op_silent:1; 369 unsigned int op_active:1; 370 unsigned int op_bytes; 371 unsigned int op_nents; 372 unsigned int op_count; 373 struct scatterlist *op_sg; 374 struct rds_notifier *op_notifier; 375 376 struct rds_mr *op_rdma_mr; 377 } rdma; 378 struct rm_data_op { 379 unsigned int op_active:1; 380 unsigned int op_nents; 381 unsigned int op_count; 382 unsigned int op_dmasg; 383 unsigned int op_dmaoff; 384 struct scatterlist *op_sg; 385 } data; 386 }; 387 }; 388 389 /* 390 * The RDS notifier is used (optionally) to tell the application about 391 * completed RDMA operations. Rather than keeping the whole rds message 392 * around on the queue, we allocate a small notifier that is put on the 393 * socket's notifier_list. Notifications are delivered to the application 394 * through control messages. 395 */ 396 struct rds_notifier { 397 struct list_head n_list; 398 uint64_t n_user_token; 399 int n_status; 400 }; 401 402 /** 403 * struct rds_transport - transport specific behavioural hooks 404 * 405 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send 406 * part of a message. The caller serializes on the send_sem so this 407 * doesn't need to be reentrant for a given conn. The header must be 408 * sent before the data payload. .xmit must be prepared to send a 409 * message with no data payload. .xmit should return the number of 410 * bytes that were sent down the connection, including header bytes. 411 * Returning 0 tells the caller that it doesn't need to perform any 412 * additional work now. This is usually the case when the transport has 413 * filled the sending queue for its connection and will handle 414 * triggering the rds thread to continue the send when space becomes 415 * available. Returning -EAGAIN tells the caller to retry the send 416 * immediately. Returning -ENOMEM tells the caller to retry the send at 417 * some point in the future. 418 * 419 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once 420 * it returns the connection can not call rds_recv_incoming(). 421 * This will only be called once after conn_connect returns 422 * non-zero success and will The caller serializes this with 423 * the send and connecting paths (xmit_* and conn_*). The 424 * transport is responsible for other serialization, including 425 * rds_recv_incoming(). This is called in process context but 426 * should try hard not to block. 427 */ 428 429 struct rds_transport { 430 char t_name[TRANSNAMSIZ]; 431 struct list_head t_item; 432 struct module *t_owner; 433 unsigned int t_prefer_loopback:1; 434 unsigned int t_type; 435 436 int (*laddr_check)(struct net *net, __be32 addr); 437 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); 438 void (*conn_free)(void *data); 439 int (*conn_connect)(struct rds_connection *conn); 440 void (*conn_shutdown)(struct rds_connection *conn); 441 void (*xmit_prepare)(struct rds_connection *conn); 442 void (*xmit_complete)(struct rds_connection *conn); 443 int (*xmit)(struct rds_connection *conn, struct rds_message *rm, 444 unsigned int hdr_off, unsigned int sg, unsigned int off); 445 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); 446 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); 447 int (*recv)(struct rds_connection *conn); 448 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); 449 void (*inc_free)(struct rds_incoming *inc); 450 451 int (*cm_handle_connect)(struct rdma_cm_id *cm_id, 452 struct rdma_cm_event *event); 453 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); 454 void (*cm_connect_complete)(struct rds_connection *conn, 455 struct rdma_cm_event *event); 456 457 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, 458 unsigned int avail); 459 void (*exit)(void); 460 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, 461 struct rds_sock *rs, u32 *key_ret); 462 void (*sync_mr)(void *trans_private, int direction); 463 void (*free_mr)(void *trans_private, int invalidate); 464 void (*flush_mrs)(void); 465 }; 466 467 struct rds_sock { 468 struct sock rs_sk; 469 470 u64 rs_user_addr; 471 u64 rs_user_bytes; 472 473 /* 474 * bound_addr used for both incoming and outgoing, no INADDR_ANY 475 * support. 476 */ 477 struct hlist_node rs_bound_node; 478 __be32 rs_bound_addr; 479 __be32 rs_conn_addr; 480 __be16 rs_bound_port; 481 __be16 rs_conn_port; 482 struct rds_transport *rs_transport; 483 484 /* 485 * rds_sendmsg caches the conn it used the last time around. 486 * This helps avoid costly lookups. 487 */ 488 struct rds_connection *rs_conn; 489 490 /* flag indicating we were congested or not */ 491 int rs_congested; 492 /* seen congestion (ENOBUFS) when sending? */ 493 int rs_seen_congestion; 494 495 /* rs_lock protects all these adjacent members before the newline */ 496 spinlock_t rs_lock; 497 struct list_head rs_send_queue; 498 u32 rs_snd_bytes; 499 int rs_rcv_bytes; 500 struct list_head rs_notify_queue; /* currently used for failed RDMAs */ 501 502 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask 503 * to decide whether the application should be woken up. 504 * If not set, we use rs_cong_track to find out whether a cong map 505 * update arrived. 506 */ 507 uint64_t rs_cong_mask; 508 uint64_t rs_cong_notify; 509 struct list_head rs_cong_list; 510 unsigned long rs_cong_track; 511 512 /* 513 * rs_recv_lock protects the receive queue, and is 514 * used to serialize with rds_release. 515 */ 516 rwlock_t rs_recv_lock; 517 struct list_head rs_recv_queue; 518 519 /* just for stats reporting */ 520 struct list_head rs_item; 521 522 /* these have their own lock */ 523 spinlock_t rs_rdma_lock; 524 struct rb_root rs_rdma_keys; 525 526 /* Socket options - in case there will be more */ 527 unsigned char rs_recverr, 528 rs_cong_monitor; 529 }; 530 531 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) 532 { 533 return container_of(sk, struct rds_sock, rs_sk); 534 } 535 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) 536 { 537 return &rs->rs_sk; 538 } 539 540 /* 541 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value 542 * to account for overhead. We don't account for overhead, we just apply 543 * the number of payload bytes to the specified value. 544 */ 545 static inline int rds_sk_sndbuf(struct rds_sock *rs) 546 { 547 return rds_rs_to_sk(rs)->sk_sndbuf / 2; 548 } 549 static inline int rds_sk_rcvbuf(struct rds_sock *rs) 550 { 551 return rds_rs_to_sk(rs)->sk_rcvbuf / 2; 552 } 553 554 struct rds_statistics { 555 uint64_t s_conn_reset; 556 uint64_t s_recv_drop_bad_checksum; 557 uint64_t s_recv_drop_old_seq; 558 uint64_t s_recv_drop_no_sock; 559 uint64_t s_recv_drop_dead_sock; 560 uint64_t s_recv_deliver_raced; 561 uint64_t s_recv_delivered; 562 uint64_t s_recv_queued; 563 uint64_t s_recv_immediate_retry; 564 uint64_t s_recv_delayed_retry; 565 uint64_t s_recv_ack_required; 566 uint64_t s_recv_rdma_bytes; 567 uint64_t s_recv_ping; 568 uint64_t s_send_queue_empty; 569 uint64_t s_send_queue_full; 570 uint64_t s_send_lock_contention; 571 uint64_t s_send_lock_queue_raced; 572 uint64_t s_send_immediate_retry; 573 uint64_t s_send_delayed_retry; 574 uint64_t s_send_drop_acked; 575 uint64_t s_send_ack_required; 576 uint64_t s_send_queued; 577 uint64_t s_send_rdma; 578 uint64_t s_send_rdma_bytes; 579 uint64_t s_send_pong; 580 uint64_t s_page_remainder_hit; 581 uint64_t s_page_remainder_miss; 582 uint64_t s_copy_to_user; 583 uint64_t s_copy_from_user; 584 uint64_t s_cong_update_queued; 585 uint64_t s_cong_update_received; 586 uint64_t s_cong_send_error; 587 uint64_t s_cong_send_blocked; 588 }; 589 590 /* af_rds.c */ 591 void rds_sock_addref(struct rds_sock *rs); 592 void rds_sock_put(struct rds_sock *rs); 593 void rds_wake_sk_sleep(struct rds_sock *rs); 594 static inline void __rds_wake_sk_sleep(struct sock *sk) 595 { 596 wait_queue_head_t *waitq = sk_sleep(sk); 597 598 if (!sock_flag(sk, SOCK_DEAD) && waitq) 599 wake_up(waitq); 600 } 601 extern wait_queue_head_t rds_poll_waitq; 602 603 604 /* bind.c */ 605 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); 606 void rds_remove_bound(struct rds_sock *rs); 607 struct rds_sock *rds_find_bound(__be32 addr, __be16 port); 608 609 /* cong.c */ 610 int rds_cong_get_maps(struct rds_connection *conn); 611 void rds_cong_add_conn(struct rds_connection *conn); 612 void rds_cong_remove_conn(struct rds_connection *conn); 613 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); 614 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); 615 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); 616 void rds_cong_queue_updates(struct rds_cong_map *map); 617 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); 618 int rds_cong_updated_since(unsigned long *recent); 619 void rds_cong_add_socket(struct rds_sock *); 620 void rds_cong_remove_socket(struct rds_sock *); 621 void rds_cong_exit(void); 622 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); 623 624 /* conn.c */ 625 int rds_conn_init(void); 626 void rds_conn_exit(void); 627 struct rds_connection *rds_conn_create(struct net *net, 628 __be32 laddr, __be32 faddr, 629 struct rds_transport *trans, gfp_t gfp); 630 struct rds_connection *rds_conn_create_outgoing(struct net *net, 631 __be32 laddr, __be32 faddr, 632 struct rds_transport *trans, gfp_t gfp); 633 void rds_conn_shutdown(struct rds_connection *conn); 634 void rds_conn_destroy(struct rds_connection *conn); 635 void rds_conn_drop(struct rds_connection *conn); 636 void rds_conn_connect_if_down(struct rds_connection *conn); 637 void rds_for_each_conn_info(struct socket *sock, unsigned int len, 638 struct rds_info_iterator *iter, 639 struct rds_info_lengths *lens, 640 int (*visitor)(struct rds_connection *, void *), 641 size_t item_len); 642 __printf(2, 3) 643 void __rds_conn_error(struct rds_connection *conn, const char *, ...); 644 #define rds_conn_error(conn, fmt...) \ 645 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt) 646 647 static inline int 648 rds_conn_transition(struct rds_connection *conn, int old, int new) 649 { 650 return atomic_cmpxchg(&conn->c_state, old, new) == old; 651 } 652 653 static inline int 654 rds_conn_state(struct rds_connection *conn) 655 { 656 return atomic_read(&conn->c_state); 657 } 658 659 static inline int 660 rds_conn_up(struct rds_connection *conn) 661 { 662 return atomic_read(&conn->c_state) == RDS_CONN_UP; 663 } 664 665 static inline int 666 rds_conn_connecting(struct rds_connection *conn) 667 { 668 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING; 669 } 670 671 /* message.c */ 672 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); 673 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); 674 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from); 675 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); 676 void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 677 __be16 dport, u64 seq); 678 int rds_message_add_extension(struct rds_header *hdr, 679 unsigned int type, const void *data, unsigned int len); 680 int rds_message_next_extension(struct rds_header *hdr, 681 unsigned int *pos, void *buf, unsigned int *buflen); 682 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); 683 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 684 void rds_message_inc_free(struct rds_incoming *inc); 685 void rds_message_addref(struct rds_message *rm); 686 void rds_message_put(struct rds_message *rm); 687 void rds_message_wait(struct rds_message *rm); 688 void rds_message_unmapped(struct rds_message *rm); 689 690 static inline void rds_message_make_checksum(struct rds_header *hdr) 691 { 692 hdr->h_csum = 0; 693 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); 694 } 695 696 static inline int rds_message_verify_checksum(const struct rds_header *hdr) 697 { 698 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; 699 } 700 701 702 /* page.c */ 703 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, 704 gfp_t gfp); 705 int rds_page_copy_user(struct page *page, unsigned long offset, 706 void __user *ptr, unsigned long bytes, 707 int to_user); 708 #define rds_page_copy_to_user(page, offset, ptr, bytes) \ 709 rds_page_copy_user(page, offset, ptr, bytes, 1) 710 #define rds_page_copy_from_user(page, offset, ptr, bytes) \ 711 rds_page_copy_user(page, offset, ptr, bytes, 0) 712 void rds_page_exit(void); 713 714 /* recv.c */ 715 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 716 __be32 saddr); 717 void rds_inc_put(struct rds_incoming *inc); 718 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, 719 struct rds_incoming *inc, gfp_t gfp); 720 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, 721 int msg_flags); 722 void rds_clear_recv_queue(struct rds_sock *rs); 723 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); 724 void rds_inc_info_copy(struct rds_incoming *inc, 725 struct rds_info_iterator *iter, 726 __be32 saddr, __be32 daddr, int flip); 727 728 /* send.c */ 729 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); 730 void rds_send_reset(struct rds_connection *conn); 731 int rds_send_xmit(struct rds_connection *conn); 732 struct sockaddr_in; 733 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); 734 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); 735 void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 736 is_acked_func is_acked); 737 int rds_send_pong(struct rds_connection *conn, __be16 dport); 738 struct rds_message *rds_send_get_message(struct rds_connection *, 739 struct rm_rdma_op *); 740 741 /* rdma.c */ 742 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); 743 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); 744 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); 745 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); 746 void rds_rdma_drop_keys(struct rds_sock *rs); 747 int rds_rdma_extra_size(struct rds_rdma_args *args); 748 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 749 struct cmsghdr *cmsg); 750 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, 751 struct cmsghdr *cmsg); 752 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 753 struct cmsghdr *cmsg); 754 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, 755 struct cmsghdr *cmsg); 756 void rds_rdma_free_op(struct rm_rdma_op *ro); 757 void rds_atomic_free_op(struct rm_atomic_op *ao); 758 void rds_rdma_send_complete(struct rds_message *rm, int wc_status); 759 void rds_atomic_send_complete(struct rds_message *rm, int wc_status); 760 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, 761 struct cmsghdr *cmsg); 762 763 void __rds_put_mr_final(struct rds_mr *mr); 764 static inline void rds_mr_put(struct rds_mr *mr) 765 { 766 if (atomic_dec_and_test(&mr->r_refcount)) 767 __rds_put_mr_final(mr); 768 } 769 770 /* stats.c */ 771 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); 772 #define rds_stats_inc_which(which, member) do { \ 773 per_cpu(which, get_cpu()).member++; \ 774 put_cpu(); \ 775 } while (0) 776 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) 777 #define rds_stats_add_which(which, member, count) do { \ 778 per_cpu(which, get_cpu()).member += count; \ 779 put_cpu(); \ 780 } while (0) 781 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) 782 int rds_stats_init(void); 783 void rds_stats_exit(void); 784 void rds_stats_info_copy(struct rds_info_iterator *iter, 785 uint64_t *values, const char *const *names, 786 size_t nr); 787 788 /* sysctl.c */ 789 int rds_sysctl_init(void); 790 void rds_sysctl_exit(void); 791 extern unsigned long rds_sysctl_sndbuf_min; 792 extern unsigned long rds_sysctl_sndbuf_default; 793 extern unsigned long rds_sysctl_sndbuf_max; 794 extern unsigned long rds_sysctl_reconnect_min_jiffies; 795 extern unsigned long rds_sysctl_reconnect_max_jiffies; 796 extern unsigned int rds_sysctl_max_unacked_packets; 797 extern unsigned int rds_sysctl_max_unacked_bytes; 798 extern unsigned int rds_sysctl_ping_enable; 799 extern unsigned long rds_sysctl_trace_flags; 800 extern unsigned int rds_sysctl_trace_level; 801 802 /* threads.c */ 803 int rds_threads_init(void); 804 void rds_threads_exit(void); 805 extern struct workqueue_struct *rds_wq; 806 void rds_queue_reconnect(struct rds_connection *conn); 807 void rds_connect_worker(struct work_struct *); 808 void rds_shutdown_worker(struct work_struct *); 809 void rds_send_worker(struct work_struct *); 810 void rds_recv_worker(struct work_struct *); 811 void rds_connect_complete(struct rds_connection *conn); 812 813 /* transport.c */ 814 int rds_trans_register(struct rds_transport *trans); 815 void rds_trans_unregister(struct rds_transport *trans); 816 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr); 817 void rds_trans_put(struct rds_transport *trans); 818 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, 819 unsigned int avail); 820 struct rds_transport *rds_trans_get(int t_type); 821 int rds_trans_init(void); 822 void rds_trans_exit(void); 823 824 #endif 825