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 #include <linux/rhashtable.h> 11 12 #include "info.h" 13 14 /* 15 * RDS Network protocol version 16 */ 17 #define RDS_PROTOCOL_3_0 0x0300 18 #define RDS_PROTOCOL_3_1 0x0301 19 #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1 20 #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8) 21 #define RDS_PROTOCOL_MINOR(v) ((v) & 255) 22 #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min) 23 24 /* 25 * XXX randomly chosen, but at least seems to be unused: 26 * # 18464-18768 Unassigned 27 * We should do better. We want a reserved port to discourage unpriv'ed 28 * userspace from listening. 29 */ 30 #define RDS_PORT 18634 31 32 #ifdef ATOMIC64_INIT 33 #define KERNEL_HAS_ATOMIC64 34 #endif 35 36 #ifdef RDS_DEBUG 37 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args) 38 #else 39 /* sigh, pr_debug() causes unused variable warnings */ 40 static inline __printf(1, 2) 41 void rdsdebug(char *fmt, ...) 42 { 43 } 44 #endif 45 46 /* XXX is there one of these somewhere? */ 47 #define ceil(x, y) \ 48 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; }) 49 50 #define RDS_FRAG_SHIFT 12 51 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT)) 52 53 #define RDS_CONG_MAP_BYTES (65536 / 8) 54 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE) 55 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8) 56 57 struct rds_cong_map { 58 struct rb_node m_rb_node; 59 __be32 m_addr; 60 wait_queue_head_t m_waitq; 61 struct list_head m_conn_list; 62 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES]; 63 }; 64 65 66 /* 67 * This is how we will track the connection state: 68 * A connection is always in one of the following 69 * states. Updates to the state are atomic and imply 70 * a memory barrier. 71 */ 72 enum { 73 RDS_CONN_DOWN = 0, 74 RDS_CONN_CONNECTING, 75 RDS_CONN_DISCONNECTING, 76 RDS_CONN_UP, 77 RDS_CONN_RESETTING, 78 RDS_CONN_ERROR, 79 }; 80 81 /* Bits for c_flags */ 82 #define RDS_LL_SEND_FULL 0 83 #define RDS_RECONNECT_PENDING 1 84 #define RDS_IN_XMIT 2 85 #define RDS_RECV_REFILL 3 86 87 /* Max number of multipaths per RDS connection. Must be a power of 2 */ 88 #define RDS_MPATH_WORKERS 8 89 #define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \ 90 (rs)->rs_hash_initval) & ((n) - 1)) 91 92 /* Per mpath connection state */ 93 struct rds_conn_path { 94 struct rds_connection *cp_conn; 95 struct rds_message *cp_xmit_rm; 96 unsigned long cp_xmit_sg; 97 unsigned int cp_xmit_hdr_off; 98 unsigned int cp_xmit_data_off; 99 unsigned int cp_xmit_atomic_sent; 100 unsigned int cp_xmit_rdma_sent; 101 unsigned int cp_xmit_data_sent; 102 103 spinlock_t cp_lock; /* protect msg queues */ 104 u64 cp_next_tx_seq; 105 struct list_head cp_send_queue; 106 struct list_head cp_retrans; 107 108 u64 cp_next_rx_seq; 109 110 void *cp_transport_data; 111 112 atomic_t cp_state; 113 unsigned long cp_send_gen; 114 unsigned long cp_flags; 115 unsigned long cp_reconnect_jiffies; 116 struct delayed_work cp_send_w; 117 struct delayed_work cp_recv_w; 118 struct delayed_work cp_conn_w; 119 struct work_struct cp_down_w; 120 struct mutex cp_cm_lock; /* protect cp_state & cm */ 121 wait_queue_head_t cp_waitq; 122 123 unsigned int cp_unacked_packets; 124 unsigned int cp_unacked_bytes; 125 unsigned int cp_outgoing:1, 126 cp_pad_to_32:31; 127 unsigned int cp_index; 128 }; 129 130 /* One rds_connection per RDS address pair */ 131 struct rds_connection { 132 struct hlist_node c_hash_node; 133 __be32 c_laddr; 134 __be32 c_faddr; 135 unsigned int c_loopback:1, 136 c_ping_triggered:1, 137 c_pad_to_32:30; 138 int c_npaths; 139 struct rds_connection *c_passive; 140 struct rds_transport *c_trans; 141 142 struct rds_cong_map *c_lcong; 143 struct rds_cong_map *c_fcong; 144 145 /* Protocol version */ 146 unsigned int c_version; 147 possible_net_t c_net; 148 149 struct list_head c_map_item; 150 unsigned long c_map_queued; 151 152 struct rds_conn_path c_path[RDS_MPATH_WORKERS]; 153 wait_queue_head_t c_hs_waitq; /* handshake waitq */ 154 155 u32 c_my_gen_num; 156 u32 c_peer_gen_num; 157 }; 158 159 static inline 160 struct net *rds_conn_net(struct rds_connection *conn) 161 { 162 return read_pnet(&conn->c_net); 163 } 164 165 static inline 166 void rds_conn_net_set(struct rds_connection *conn, struct net *net) 167 { 168 write_pnet(&conn->c_net, net); 169 } 170 171 #define RDS_FLAG_CONG_BITMAP 0x01 172 #define RDS_FLAG_ACK_REQUIRED 0x02 173 #define RDS_FLAG_RETRANSMITTED 0x04 174 #define RDS_MAX_ADV_CREDIT 255 175 176 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping 177 * probe to exchange control information before establishing a connection. 178 * Currently the control information that is exchanged is the number of 179 * supported paths. If the peer is a legacy (older kernel revision) peer, 180 * it would return a pong message without additional control information 181 * that would then alert the sender that the peer was an older rev. 182 */ 183 #define RDS_FLAG_PROBE_PORT 1 184 #define RDS_HS_PROBE(sport, dport) \ 185 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \ 186 (sport == 0 && dport == RDS_FLAG_PROBE_PORT)) 187 /* 188 * Maximum space available for extension headers. 189 */ 190 #define RDS_HEADER_EXT_SPACE 16 191 192 struct rds_header { 193 __be64 h_sequence; 194 __be64 h_ack; 195 __be32 h_len; 196 __be16 h_sport; 197 __be16 h_dport; 198 u8 h_flags; 199 u8 h_credit; 200 u8 h_padding[4]; 201 __sum16 h_csum; 202 203 u8 h_exthdr[RDS_HEADER_EXT_SPACE]; 204 }; 205 206 /* 207 * Reserved - indicates end of extensions 208 */ 209 #define RDS_EXTHDR_NONE 0 210 211 /* 212 * This extension header is included in the very 213 * first message that is sent on a new connection, 214 * and identifies the protocol level. This will help 215 * rolling updates if a future change requires breaking 216 * the protocol. 217 * NB: This is no longer true for IB, where we do a version 218 * negotiation during the connection setup phase (protocol 219 * version information is included in the RDMA CM private data). 220 */ 221 #define RDS_EXTHDR_VERSION 1 222 struct rds_ext_header_version { 223 __be32 h_version; 224 }; 225 226 /* 227 * This extension header is included in the RDS message 228 * chasing an RDMA operation. 229 */ 230 #define RDS_EXTHDR_RDMA 2 231 struct rds_ext_header_rdma { 232 __be32 h_rdma_rkey; 233 }; 234 235 /* 236 * This extension header tells the peer about the 237 * destination <R_Key,offset> of the requested RDMA 238 * operation. 239 */ 240 #define RDS_EXTHDR_RDMA_DEST 3 241 struct rds_ext_header_rdma_dest { 242 __be32 h_rdma_rkey; 243 __be32 h_rdma_offset; 244 }; 245 246 /* Extension header announcing number of paths. 247 * Implicit length = 2 bytes. 248 */ 249 #define RDS_EXTHDR_NPATHS 5 250 #define RDS_EXTHDR_GEN_NUM 6 251 252 #define __RDS_EXTHDR_MAX 16 /* for now */ 253 254 struct rds_incoming { 255 atomic_t i_refcount; 256 struct list_head i_item; 257 struct rds_connection *i_conn; 258 struct rds_conn_path *i_conn_path; 259 struct rds_header i_hdr; 260 unsigned long i_rx_jiffies; 261 __be32 i_saddr; 262 263 rds_rdma_cookie_t i_rdma_cookie; 264 struct timeval i_rx_tstamp; 265 }; 266 267 struct rds_mr { 268 struct rb_node r_rb_node; 269 atomic_t r_refcount; 270 u32 r_key; 271 272 /* A copy of the creation flags */ 273 unsigned int r_use_once:1; 274 unsigned int r_invalidate:1; 275 unsigned int r_write:1; 276 277 /* This is for RDS_MR_DEAD. 278 * It would be nice & consistent to make this part of the above 279 * bit field here, but we need to use test_and_set_bit. 280 */ 281 unsigned long r_state; 282 struct rds_sock *r_sock; /* back pointer to the socket that owns us */ 283 struct rds_transport *r_trans; 284 void *r_trans_private; 285 }; 286 287 /* Flags for mr->r_state */ 288 #define RDS_MR_DEAD 0 289 290 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) 291 { 292 return r_key | (((u64) offset) << 32); 293 } 294 295 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) 296 { 297 return cookie; 298 } 299 300 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) 301 { 302 return cookie >> 32; 303 } 304 305 /* atomic operation types */ 306 #define RDS_ATOMIC_TYPE_CSWP 0 307 #define RDS_ATOMIC_TYPE_FADD 1 308 309 /* 310 * m_sock_item and m_conn_item are on lists that are serialized under 311 * conn->c_lock. m_sock_item has additional meaning in that once it is empty 312 * the message will not be put back on the retransmit list after being sent. 313 * messages that are canceled while being sent rely on this. 314 * 315 * m_inc is used by loopback so that it can pass an incoming message straight 316 * back up into the rx path. It embeds a wire header which is also used by 317 * the send path, which is kind of awkward. 318 * 319 * m_sock_item indicates the message's presence on a socket's send or receive 320 * queue. m_rs will point to that socket. 321 * 322 * m_daddr is used by cancellation to prune messages to a given destination. 323 * 324 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock 325 * nesting. As paths iterate over messages on a sock, or conn, they must 326 * also lock the conn, or sock, to remove the message from those lists too. 327 * Testing the flag to determine if the message is still on the lists lets 328 * us avoid testing the list_head directly. That means each path can use 329 * the message's list_head to keep it on a local list while juggling locks 330 * without confusing the other path. 331 * 332 * m_ack_seq is an optional field set by transports who need a different 333 * sequence number range to invalidate. They can use this in a callback 334 * that they pass to rds_send_drop_acked() to see if each message has been 335 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't 336 * had ack_seq set yet. 337 */ 338 #define RDS_MSG_ON_SOCK 1 339 #define RDS_MSG_ON_CONN 2 340 #define RDS_MSG_HAS_ACK_SEQ 3 341 #define RDS_MSG_ACK_REQUIRED 4 342 #define RDS_MSG_RETRANSMITTED 5 343 #define RDS_MSG_MAPPED 6 344 #define RDS_MSG_PAGEVEC 7 345 #define RDS_MSG_FLUSH 8 346 347 struct rds_message { 348 atomic_t m_refcount; 349 struct list_head m_sock_item; 350 struct list_head m_conn_item; 351 struct rds_incoming m_inc; 352 u64 m_ack_seq; 353 __be32 m_daddr; 354 unsigned long m_flags; 355 356 /* Never access m_rs without holding m_rs_lock. 357 * Lock nesting is 358 * rm->m_rs_lock 359 * -> rs->rs_lock 360 */ 361 spinlock_t m_rs_lock; 362 wait_queue_head_t m_flush_wait; 363 364 struct rds_sock *m_rs; 365 366 /* cookie to send to remote, in rds header */ 367 rds_rdma_cookie_t m_rdma_cookie; 368 369 unsigned int m_used_sgs; 370 unsigned int m_total_sgs; 371 372 void *m_final_op; 373 374 struct { 375 struct rm_atomic_op { 376 int op_type; 377 union { 378 struct { 379 uint64_t compare; 380 uint64_t swap; 381 uint64_t compare_mask; 382 uint64_t swap_mask; 383 } op_m_cswp; 384 struct { 385 uint64_t add; 386 uint64_t nocarry_mask; 387 } op_m_fadd; 388 }; 389 390 u32 op_rkey; 391 u64 op_remote_addr; 392 unsigned int op_notify:1; 393 unsigned int op_recverr:1; 394 unsigned int op_mapped:1; 395 unsigned int op_silent:1; 396 unsigned int op_active:1; 397 struct scatterlist *op_sg; 398 struct rds_notifier *op_notifier; 399 400 struct rds_mr *op_rdma_mr; 401 } atomic; 402 struct rm_rdma_op { 403 u32 op_rkey; 404 u64 op_remote_addr; 405 unsigned int op_write:1; 406 unsigned int op_fence:1; 407 unsigned int op_notify:1; 408 unsigned int op_recverr:1; 409 unsigned int op_mapped:1; 410 unsigned int op_silent:1; 411 unsigned int op_active:1; 412 unsigned int op_bytes; 413 unsigned int op_nents; 414 unsigned int op_count; 415 struct scatterlist *op_sg; 416 struct rds_notifier *op_notifier; 417 418 struct rds_mr *op_rdma_mr; 419 } rdma; 420 struct rm_data_op { 421 unsigned int op_active:1; 422 unsigned int op_nents; 423 unsigned int op_count; 424 unsigned int op_dmasg; 425 unsigned int op_dmaoff; 426 struct scatterlist *op_sg; 427 } data; 428 }; 429 }; 430 431 /* 432 * The RDS notifier is used (optionally) to tell the application about 433 * completed RDMA operations. Rather than keeping the whole rds message 434 * around on the queue, we allocate a small notifier that is put on the 435 * socket's notifier_list. Notifications are delivered to the application 436 * through control messages. 437 */ 438 struct rds_notifier { 439 struct list_head n_list; 440 uint64_t n_user_token; 441 int n_status; 442 }; 443 444 /** 445 * struct rds_transport - transport specific behavioural hooks 446 * 447 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send 448 * part of a message. The caller serializes on the send_sem so this 449 * doesn't need to be reentrant for a given conn. The header must be 450 * sent before the data payload. .xmit must be prepared to send a 451 * message with no data payload. .xmit should return the number of 452 * bytes that were sent down the connection, including header bytes. 453 * Returning 0 tells the caller that it doesn't need to perform any 454 * additional work now. This is usually the case when the transport has 455 * filled the sending queue for its connection and will handle 456 * triggering the rds thread to continue the send when space becomes 457 * available. Returning -EAGAIN tells the caller to retry the send 458 * immediately. Returning -ENOMEM tells the caller to retry the send at 459 * some point in the future. 460 * 461 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once 462 * it returns the connection can not call rds_recv_incoming(). 463 * This will only be called once after conn_connect returns 464 * non-zero success and will The caller serializes this with 465 * the send and connecting paths (xmit_* and conn_*). The 466 * transport is responsible for other serialization, including 467 * rds_recv_incoming(). This is called in process context but 468 * should try hard not to block. 469 */ 470 471 struct rds_transport { 472 char t_name[TRANSNAMSIZ]; 473 struct list_head t_item; 474 struct module *t_owner; 475 unsigned int t_prefer_loopback:1, 476 t_mp_capable:1; 477 unsigned int t_type; 478 479 int (*laddr_check)(struct net *net, __be32 addr); 480 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); 481 void (*conn_free)(void *data); 482 int (*conn_path_connect)(struct rds_conn_path *cp); 483 void (*conn_path_shutdown)(struct rds_conn_path *conn); 484 void (*xmit_path_prepare)(struct rds_conn_path *cp); 485 void (*xmit_path_complete)(struct rds_conn_path *cp); 486 int (*xmit)(struct rds_connection *conn, struct rds_message *rm, 487 unsigned int hdr_off, unsigned int sg, unsigned int off); 488 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); 489 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); 490 int (*recv_path)(struct rds_conn_path *cp); 491 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); 492 void (*inc_free)(struct rds_incoming *inc); 493 494 int (*cm_handle_connect)(struct rdma_cm_id *cm_id, 495 struct rdma_cm_event *event); 496 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); 497 void (*cm_connect_complete)(struct rds_connection *conn, 498 struct rdma_cm_event *event); 499 500 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, 501 unsigned int avail); 502 void (*exit)(void); 503 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, 504 struct rds_sock *rs, u32 *key_ret); 505 void (*sync_mr)(void *trans_private, int direction); 506 void (*free_mr)(void *trans_private, int invalidate); 507 void (*flush_mrs)(void); 508 }; 509 510 struct rds_sock { 511 struct sock rs_sk; 512 513 u64 rs_user_addr; 514 u64 rs_user_bytes; 515 516 /* 517 * bound_addr used for both incoming and outgoing, no INADDR_ANY 518 * support. 519 */ 520 struct rhash_head rs_bound_node; 521 u64 rs_bound_key; 522 __be32 rs_bound_addr; 523 __be32 rs_conn_addr; 524 __be16 rs_bound_port; 525 __be16 rs_conn_port; 526 struct rds_transport *rs_transport; 527 528 /* 529 * rds_sendmsg caches the conn it used the last time around. 530 * This helps avoid costly lookups. 531 */ 532 struct rds_connection *rs_conn; 533 534 /* flag indicating we were congested or not */ 535 int rs_congested; 536 /* seen congestion (ENOBUFS) when sending? */ 537 int rs_seen_congestion; 538 539 /* rs_lock protects all these adjacent members before the newline */ 540 spinlock_t rs_lock; 541 struct list_head rs_send_queue; 542 u32 rs_snd_bytes; 543 int rs_rcv_bytes; 544 struct list_head rs_notify_queue; /* currently used for failed RDMAs */ 545 546 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask 547 * to decide whether the application should be woken up. 548 * If not set, we use rs_cong_track to find out whether a cong map 549 * update arrived. 550 */ 551 uint64_t rs_cong_mask; 552 uint64_t rs_cong_notify; 553 struct list_head rs_cong_list; 554 unsigned long rs_cong_track; 555 556 /* 557 * rs_recv_lock protects the receive queue, and is 558 * used to serialize with rds_release. 559 */ 560 rwlock_t rs_recv_lock; 561 struct list_head rs_recv_queue; 562 563 /* just for stats reporting */ 564 struct list_head rs_item; 565 566 /* these have their own lock */ 567 spinlock_t rs_rdma_lock; 568 struct rb_root rs_rdma_keys; 569 570 /* Socket options - in case there will be more */ 571 unsigned char rs_recverr, 572 rs_cong_monitor; 573 u32 rs_hash_initval; 574 }; 575 576 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) 577 { 578 return container_of(sk, struct rds_sock, rs_sk); 579 } 580 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) 581 { 582 return &rs->rs_sk; 583 } 584 585 /* 586 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value 587 * to account for overhead. We don't account for overhead, we just apply 588 * the number of payload bytes to the specified value. 589 */ 590 static inline int rds_sk_sndbuf(struct rds_sock *rs) 591 { 592 return rds_rs_to_sk(rs)->sk_sndbuf / 2; 593 } 594 static inline int rds_sk_rcvbuf(struct rds_sock *rs) 595 { 596 return rds_rs_to_sk(rs)->sk_rcvbuf / 2; 597 } 598 599 struct rds_statistics { 600 uint64_t s_conn_reset; 601 uint64_t s_recv_drop_bad_checksum; 602 uint64_t s_recv_drop_old_seq; 603 uint64_t s_recv_drop_no_sock; 604 uint64_t s_recv_drop_dead_sock; 605 uint64_t s_recv_deliver_raced; 606 uint64_t s_recv_delivered; 607 uint64_t s_recv_queued; 608 uint64_t s_recv_immediate_retry; 609 uint64_t s_recv_delayed_retry; 610 uint64_t s_recv_ack_required; 611 uint64_t s_recv_rdma_bytes; 612 uint64_t s_recv_ping; 613 uint64_t s_send_queue_empty; 614 uint64_t s_send_queue_full; 615 uint64_t s_send_lock_contention; 616 uint64_t s_send_lock_queue_raced; 617 uint64_t s_send_immediate_retry; 618 uint64_t s_send_delayed_retry; 619 uint64_t s_send_drop_acked; 620 uint64_t s_send_ack_required; 621 uint64_t s_send_queued; 622 uint64_t s_send_rdma; 623 uint64_t s_send_rdma_bytes; 624 uint64_t s_send_pong; 625 uint64_t s_page_remainder_hit; 626 uint64_t s_page_remainder_miss; 627 uint64_t s_copy_to_user; 628 uint64_t s_copy_from_user; 629 uint64_t s_cong_update_queued; 630 uint64_t s_cong_update_received; 631 uint64_t s_cong_send_error; 632 uint64_t s_cong_send_blocked; 633 }; 634 635 /* af_rds.c */ 636 void rds_sock_addref(struct rds_sock *rs); 637 void rds_sock_put(struct rds_sock *rs); 638 void rds_wake_sk_sleep(struct rds_sock *rs); 639 static inline void __rds_wake_sk_sleep(struct sock *sk) 640 { 641 wait_queue_head_t *waitq = sk_sleep(sk); 642 643 if (!sock_flag(sk, SOCK_DEAD) && waitq) 644 wake_up(waitq); 645 } 646 extern wait_queue_head_t rds_poll_waitq; 647 648 649 /* bind.c */ 650 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); 651 void rds_remove_bound(struct rds_sock *rs); 652 struct rds_sock *rds_find_bound(__be32 addr, __be16 port); 653 int rds_bind_lock_init(void); 654 void rds_bind_lock_destroy(void); 655 656 /* cong.c */ 657 int rds_cong_get_maps(struct rds_connection *conn); 658 void rds_cong_add_conn(struct rds_connection *conn); 659 void rds_cong_remove_conn(struct rds_connection *conn); 660 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); 661 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); 662 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); 663 void rds_cong_queue_updates(struct rds_cong_map *map); 664 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); 665 int rds_cong_updated_since(unsigned long *recent); 666 void rds_cong_add_socket(struct rds_sock *); 667 void rds_cong_remove_socket(struct rds_sock *); 668 void rds_cong_exit(void); 669 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); 670 671 /* conn.c */ 672 extern u32 rds_gen_num; 673 int rds_conn_init(void); 674 void rds_conn_exit(void); 675 struct rds_connection *rds_conn_create(struct net *net, 676 __be32 laddr, __be32 faddr, 677 struct rds_transport *trans, gfp_t gfp); 678 struct rds_connection *rds_conn_create_outgoing(struct net *net, 679 __be32 laddr, __be32 faddr, 680 struct rds_transport *trans, gfp_t gfp); 681 void rds_conn_shutdown(struct rds_conn_path *cpath); 682 void rds_conn_destroy(struct rds_connection *conn); 683 void rds_conn_drop(struct rds_connection *conn); 684 void rds_conn_path_drop(struct rds_conn_path *cpath); 685 void rds_conn_connect_if_down(struct rds_connection *conn); 686 void rds_conn_path_connect_if_down(struct rds_conn_path *cp); 687 void rds_for_each_conn_info(struct socket *sock, unsigned int len, 688 struct rds_info_iterator *iter, 689 struct rds_info_lengths *lens, 690 int (*visitor)(struct rds_connection *, void *), 691 size_t item_len); 692 693 __printf(2, 3) 694 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...); 695 #define rds_conn_path_error(cp, fmt...) \ 696 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt) 697 698 static inline int 699 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new) 700 { 701 return atomic_cmpxchg(&cp->cp_state, old, new) == old; 702 } 703 704 static inline int 705 rds_conn_transition(struct rds_connection *conn, int old, int new) 706 { 707 WARN_ON(conn->c_trans->t_mp_capable); 708 return rds_conn_path_transition(&conn->c_path[0], old, new); 709 } 710 711 static inline int 712 rds_conn_path_state(struct rds_conn_path *cp) 713 { 714 return atomic_read(&cp->cp_state); 715 } 716 717 static inline int 718 rds_conn_state(struct rds_connection *conn) 719 { 720 WARN_ON(conn->c_trans->t_mp_capable); 721 return rds_conn_path_state(&conn->c_path[0]); 722 } 723 724 static inline int 725 rds_conn_path_up(struct rds_conn_path *cp) 726 { 727 return atomic_read(&cp->cp_state) == RDS_CONN_UP; 728 } 729 730 static inline int 731 rds_conn_up(struct rds_connection *conn) 732 { 733 WARN_ON(conn->c_trans->t_mp_capable); 734 return rds_conn_path_up(&conn->c_path[0]); 735 } 736 737 static inline int 738 rds_conn_path_connecting(struct rds_conn_path *cp) 739 { 740 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING; 741 } 742 743 static inline int 744 rds_conn_connecting(struct rds_connection *conn) 745 { 746 WARN_ON(conn->c_trans->t_mp_capable); 747 return rds_conn_path_connecting(&conn->c_path[0]); 748 } 749 750 /* message.c */ 751 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); 752 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); 753 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from); 754 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); 755 void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 756 __be16 dport, u64 seq); 757 int rds_message_add_extension(struct rds_header *hdr, 758 unsigned int type, const void *data, unsigned int len); 759 int rds_message_next_extension(struct rds_header *hdr, 760 unsigned int *pos, void *buf, unsigned int *buflen); 761 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); 762 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 763 void rds_message_inc_free(struct rds_incoming *inc); 764 void rds_message_addref(struct rds_message *rm); 765 void rds_message_put(struct rds_message *rm); 766 void rds_message_wait(struct rds_message *rm); 767 void rds_message_unmapped(struct rds_message *rm); 768 769 static inline void rds_message_make_checksum(struct rds_header *hdr) 770 { 771 hdr->h_csum = 0; 772 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); 773 } 774 775 static inline int rds_message_verify_checksum(const struct rds_header *hdr) 776 { 777 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; 778 } 779 780 781 /* page.c */ 782 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, 783 gfp_t gfp); 784 int rds_page_copy_user(struct page *page, unsigned long offset, 785 void __user *ptr, unsigned long bytes, 786 int to_user); 787 #define rds_page_copy_to_user(page, offset, ptr, bytes) \ 788 rds_page_copy_user(page, offset, ptr, bytes, 1) 789 #define rds_page_copy_from_user(page, offset, ptr, bytes) \ 790 rds_page_copy_user(page, offset, ptr, bytes, 0) 791 void rds_page_exit(void); 792 793 /* recv.c */ 794 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 795 __be32 saddr); 796 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn, 797 __be32 saddr); 798 void rds_inc_put(struct rds_incoming *inc); 799 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, 800 struct rds_incoming *inc, gfp_t gfp); 801 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, 802 int msg_flags); 803 void rds_clear_recv_queue(struct rds_sock *rs); 804 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); 805 void rds_inc_info_copy(struct rds_incoming *inc, 806 struct rds_info_iterator *iter, 807 __be32 saddr, __be32 daddr, int flip); 808 809 /* send.c */ 810 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); 811 void rds_send_path_reset(struct rds_conn_path *conn); 812 int rds_send_xmit(struct rds_conn_path *cp); 813 struct sockaddr_in; 814 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); 815 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); 816 void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 817 is_acked_func is_acked); 818 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, 819 is_acked_func is_acked); 820 int rds_send_pong(struct rds_conn_path *cp, __be16 dport); 821 822 /* rdma.c */ 823 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); 824 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); 825 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); 826 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); 827 void rds_rdma_drop_keys(struct rds_sock *rs); 828 int rds_rdma_extra_size(struct rds_rdma_args *args); 829 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 830 struct cmsghdr *cmsg); 831 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, 832 struct cmsghdr *cmsg); 833 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 834 struct cmsghdr *cmsg); 835 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, 836 struct cmsghdr *cmsg); 837 void rds_rdma_free_op(struct rm_rdma_op *ro); 838 void rds_atomic_free_op(struct rm_atomic_op *ao); 839 void rds_rdma_send_complete(struct rds_message *rm, int wc_status); 840 void rds_atomic_send_complete(struct rds_message *rm, int wc_status); 841 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, 842 struct cmsghdr *cmsg); 843 844 void __rds_put_mr_final(struct rds_mr *mr); 845 static inline void rds_mr_put(struct rds_mr *mr) 846 { 847 if (atomic_dec_and_test(&mr->r_refcount)) 848 __rds_put_mr_final(mr); 849 } 850 851 /* stats.c */ 852 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); 853 #define rds_stats_inc_which(which, member) do { \ 854 per_cpu(which, get_cpu()).member++; \ 855 put_cpu(); \ 856 } while (0) 857 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) 858 #define rds_stats_add_which(which, member, count) do { \ 859 per_cpu(which, get_cpu()).member += count; \ 860 put_cpu(); \ 861 } while (0) 862 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) 863 int rds_stats_init(void); 864 void rds_stats_exit(void); 865 void rds_stats_info_copy(struct rds_info_iterator *iter, 866 uint64_t *values, const char *const *names, 867 size_t nr); 868 869 /* sysctl.c */ 870 int rds_sysctl_init(void); 871 void rds_sysctl_exit(void); 872 extern unsigned long rds_sysctl_sndbuf_min; 873 extern unsigned long rds_sysctl_sndbuf_default; 874 extern unsigned long rds_sysctl_sndbuf_max; 875 extern unsigned long rds_sysctl_reconnect_min_jiffies; 876 extern unsigned long rds_sysctl_reconnect_max_jiffies; 877 extern unsigned int rds_sysctl_max_unacked_packets; 878 extern unsigned int rds_sysctl_max_unacked_bytes; 879 extern unsigned int rds_sysctl_ping_enable; 880 extern unsigned long rds_sysctl_trace_flags; 881 extern unsigned int rds_sysctl_trace_level; 882 883 /* threads.c */ 884 int rds_threads_init(void); 885 void rds_threads_exit(void); 886 extern struct workqueue_struct *rds_wq; 887 void rds_queue_reconnect(struct rds_conn_path *cp); 888 void rds_connect_worker(struct work_struct *); 889 void rds_shutdown_worker(struct work_struct *); 890 void rds_send_worker(struct work_struct *); 891 void rds_recv_worker(struct work_struct *); 892 void rds_connect_path_complete(struct rds_conn_path *conn, int curr); 893 void rds_connect_complete(struct rds_connection *conn); 894 895 /* transport.c */ 896 int rds_trans_register(struct rds_transport *trans); 897 void rds_trans_unregister(struct rds_transport *trans); 898 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr); 899 void rds_trans_put(struct rds_transport *trans); 900 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, 901 unsigned int avail); 902 struct rds_transport *rds_trans_get(int t_type); 903 int rds_trans_init(void); 904 void rds_trans_exit(void); 905 906 #endif 907