1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _RDS_RDS_H 3 #define _RDS_RDS_H 4 5 #include <net/sock.h> 6 #include <linux/scatterlist.h> 7 #include <linux/highmem.h> 8 #include <rdma/rdma_cm.h> 9 #include <linux/mutex.h> 10 #include <linux/rds.h> 11 #include <linux/rhashtable.h> 12 #include <linux/refcount.h> 13 14 #include "info.h" 15 16 /* 17 * RDS Network protocol version 18 */ 19 #define RDS_PROTOCOL_3_0 0x0300 20 #define RDS_PROTOCOL_3_1 0x0301 21 #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1 22 #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8) 23 #define RDS_PROTOCOL_MINOR(v) ((v) & 255) 24 #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min) 25 26 /* 27 * XXX randomly chosen, but at least seems to be unused: 28 * # 18464-18768 Unassigned 29 * We should do better. We want a reserved port to discourage unpriv'ed 30 * userspace from listening. 31 */ 32 #define RDS_PORT 18634 33 34 #ifdef ATOMIC64_INIT 35 #define KERNEL_HAS_ATOMIC64 36 #endif 37 38 #ifdef RDS_DEBUG 39 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args) 40 #else 41 /* sigh, pr_debug() causes unused variable warnings */ 42 static inline __printf(1, 2) 43 void rdsdebug(char *fmt, ...) 44 { 45 } 46 #endif 47 48 /* XXX is there one of these somewhere? */ 49 #define ceil(x, y) \ 50 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; }) 51 52 #define RDS_FRAG_SHIFT 12 53 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT)) 54 55 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */ 56 #define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20)) 57 58 #define RDS_CONG_MAP_BYTES (65536 / 8) 59 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE) 60 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8) 61 62 struct rds_cong_map { 63 struct rb_node m_rb_node; 64 __be32 m_addr; 65 wait_queue_head_t m_waitq; 66 struct list_head m_conn_list; 67 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES]; 68 }; 69 70 71 /* 72 * This is how we will track the connection state: 73 * A connection is always in one of the following 74 * states. Updates to the state are atomic and imply 75 * a memory barrier. 76 */ 77 enum { 78 RDS_CONN_DOWN = 0, 79 RDS_CONN_CONNECTING, 80 RDS_CONN_DISCONNECTING, 81 RDS_CONN_UP, 82 RDS_CONN_RESETTING, 83 RDS_CONN_ERROR, 84 }; 85 86 /* Bits for c_flags */ 87 #define RDS_LL_SEND_FULL 0 88 #define RDS_RECONNECT_PENDING 1 89 #define RDS_IN_XMIT 2 90 #define RDS_RECV_REFILL 3 91 #define RDS_DESTROY_PENDING 4 92 93 /* Max number of multipaths per RDS connection. Must be a power of 2 */ 94 #define RDS_MPATH_WORKERS 8 95 #define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \ 96 (rs)->rs_hash_initval) & ((n) - 1)) 97 98 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr)) 99 100 /* Per mpath connection state */ 101 struct rds_conn_path { 102 struct rds_connection *cp_conn; 103 struct rds_message *cp_xmit_rm; 104 unsigned long cp_xmit_sg; 105 unsigned int cp_xmit_hdr_off; 106 unsigned int cp_xmit_data_off; 107 unsigned int cp_xmit_atomic_sent; 108 unsigned int cp_xmit_rdma_sent; 109 unsigned int cp_xmit_data_sent; 110 111 spinlock_t cp_lock; /* protect msg queues */ 112 u64 cp_next_tx_seq; 113 struct list_head cp_send_queue; 114 struct list_head cp_retrans; 115 116 u64 cp_next_rx_seq; 117 118 void *cp_transport_data; 119 120 atomic_t cp_state; 121 unsigned long cp_send_gen; 122 unsigned long cp_flags; 123 unsigned long cp_reconnect_jiffies; 124 struct delayed_work cp_send_w; 125 struct delayed_work cp_recv_w; 126 struct delayed_work cp_conn_w; 127 struct work_struct cp_down_w; 128 struct mutex cp_cm_lock; /* protect cp_state & cm */ 129 wait_queue_head_t cp_waitq; 130 131 unsigned int cp_unacked_packets; 132 unsigned int cp_unacked_bytes; 133 unsigned int cp_index; 134 }; 135 136 /* One rds_connection per RDS address pair */ 137 struct rds_connection { 138 struct hlist_node c_hash_node; 139 __be32 c_laddr; 140 __be32 c_faddr; 141 unsigned int c_loopback:1, 142 c_ping_triggered:1, 143 c_pad_to_32:30; 144 int c_npaths; 145 struct rds_connection *c_passive; 146 struct rds_transport *c_trans; 147 148 struct rds_cong_map *c_lcong; 149 struct rds_cong_map *c_fcong; 150 151 /* Protocol version */ 152 unsigned int c_version; 153 possible_net_t c_net; 154 155 struct list_head c_map_item; 156 unsigned long c_map_queued; 157 158 struct rds_conn_path *c_path; 159 wait_queue_head_t c_hs_waitq; /* handshake waitq */ 160 161 u32 c_my_gen_num; 162 u32 c_peer_gen_num; 163 }; 164 165 static inline 166 struct net *rds_conn_net(struct rds_connection *conn) 167 { 168 return read_pnet(&conn->c_net); 169 } 170 171 static inline 172 void rds_conn_net_set(struct rds_connection *conn, struct net *net) 173 { 174 write_pnet(&conn->c_net, net); 175 } 176 177 #define RDS_FLAG_CONG_BITMAP 0x01 178 #define RDS_FLAG_ACK_REQUIRED 0x02 179 #define RDS_FLAG_RETRANSMITTED 0x04 180 #define RDS_MAX_ADV_CREDIT 255 181 182 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping 183 * probe to exchange control information before establishing a connection. 184 * Currently the control information that is exchanged is the number of 185 * supported paths. If the peer is a legacy (older kernel revision) peer, 186 * it would return a pong message without additional control information 187 * that would then alert the sender that the peer was an older rev. 188 */ 189 #define RDS_FLAG_PROBE_PORT 1 190 #define RDS_HS_PROBE(sport, dport) \ 191 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \ 192 (sport == 0 && dport == RDS_FLAG_PROBE_PORT)) 193 /* 194 * Maximum space available for extension headers. 195 */ 196 #define RDS_HEADER_EXT_SPACE 16 197 198 struct rds_header { 199 __be64 h_sequence; 200 __be64 h_ack; 201 __be32 h_len; 202 __be16 h_sport; 203 __be16 h_dport; 204 u8 h_flags; 205 u8 h_credit; 206 u8 h_padding[4]; 207 __sum16 h_csum; 208 209 u8 h_exthdr[RDS_HEADER_EXT_SPACE]; 210 }; 211 212 /* 213 * Reserved - indicates end of extensions 214 */ 215 #define RDS_EXTHDR_NONE 0 216 217 /* 218 * This extension header is included in the very 219 * first message that is sent on a new connection, 220 * and identifies the protocol level. This will help 221 * rolling updates if a future change requires breaking 222 * the protocol. 223 * NB: This is no longer true for IB, where we do a version 224 * negotiation during the connection setup phase (protocol 225 * version information is included in the RDMA CM private data). 226 */ 227 #define RDS_EXTHDR_VERSION 1 228 struct rds_ext_header_version { 229 __be32 h_version; 230 }; 231 232 /* 233 * This extension header is included in the RDS message 234 * chasing an RDMA operation. 235 */ 236 #define RDS_EXTHDR_RDMA 2 237 struct rds_ext_header_rdma { 238 __be32 h_rdma_rkey; 239 }; 240 241 /* 242 * This extension header tells the peer about the 243 * destination <R_Key,offset> of the requested RDMA 244 * operation. 245 */ 246 #define RDS_EXTHDR_RDMA_DEST 3 247 struct rds_ext_header_rdma_dest { 248 __be32 h_rdma_rkey; 249 __be32 h_rdma_offset; 250 }; 251 252 /* Extension header announcing number of paths. 253 * Implicit length = 2 bytes. 254 */ 255 #define RDS_EXTHDR_NPATHS 5 256 #define RDS_EXTHDR_GEN_NUM 6 257 258 #define __RDS_EXTHDR_MAX 16 /* for now */ 259 #define RDS_RX_MAX_TRACES (RDS_MSG_RX_DGRAM_TRACE_MAX + 1) 260 #define RDS_MSG_RX_HDR 0 261 #define RDS_MSG_RX_START 1 262 #define RDS_MSG_RX_END 2 263 #define RDS_MSG_RX_CMSG 3 264 265 struct rds_incoming { 266 refcount_t i_refcount; 267 struct list_head i_item; 268 struct rds_connection *i_conn; 269 struct rds_conn_path *i_conn_path; 270 struct rds_header i_hdr; 271 unsigned long i_rx_jiffies; 272 __be32 i_saddr; 273 274 rds_rdma_cookie_t i_rdma_cookie; 275 struct timeval i_rx_tstamp; 276 u64 i_rx_lat_trace[RDS_RX_MAX_TRACES]; 277 }; 278 279 struct rds_mr { 280 struct rb_node r_rb_node; 281 refcount_t r_refcount; 282 u32 r_key; 283 284 /* A copy of the creation flags */ 285 unsigned int r_use_once:1; 286 unsigned int r_invalidate:1; 287 unsigned int r_write:1; 288 289 /* This is for RDS_MR_DEAD. 290 * It would be nice & consistent to make this part of the above 291 * bit field here, but we need to use test_and_set_bit. 292 */ 293 unsigned long r_state; 294 struct rds_sock *r_sock; /* back pointer to the socket that owns us */ 295 struct rds_transport *r_trans; 296 void *r_trans_private; 297 }; 298 299 /* Flags for mr->r_state */ 300 #define RDS_MR_DEAD 0 301 302 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) 303 { 304 return r_key | (((u64) offset) << 32); 305 } 306 307 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) 308 { 309 return cookie; 310 } 311 312 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) 313 { 314 return cookie >> 32; 315 } 316 317 /* atomic operation types */ 318 #define RDS_ATOMIC_TYPE_CSWP 0 319 #define RDS_ATOMIC_TYPE_FADD 1 320 321 /* 322 * m_sock_item and m_conn_item are on lists that are serialized under 323 * conn->c_lock. m_sock_item has additional meaning in that once it is empty 324 * the message will not be put back on the retransmit list after being sent. 325 * messages that are canceled while being sent rely on this. 326 * 327 * m_inc is used by loopback so that it can pass an incoming message straight 328 * back up into the rx path. It embeds a wire header which is also used by 329 * the send path, which is kind of awkward. 330 * 331 * m_sock_item indicates the message's presence on a socket's send or receive 332 * queue. m_rs will point to that socket. 333 * 334 * m_daddr is used by cancellation to prune messages to a given destination. 335 * 336 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock 337 * nesting. As paths iterate over messages on a sock, or conn, they must 338 * also lock the conn, or sock, to remove the message from those lists too. 339 * Testing the flag to determine if the message is still on the lists lets 340 * us avoid testing the list_head directly. That means each path can use 341 * the message's list_head to keep it on a local list while juggling locks 342 * without confusing the other path. 343 * 344 * m_ack_seq is an optional field set by transports who need a different 345 * sequence number range to invalidate. They can use this in a callback 346 * that they pass to rds_send_drop_acked() to see if each message has been 347 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't 348 * had ack_seq set yet. 349 */ 350 #define RDS_MSG_ON_SOCK 1 351 #define RDS_MSG_ON_CONN 2 352 #define RDS_MSG_HAS_ACK_SEQ 3 353 #define RDS_MSG_ACK_REQUIRED 4 354 #define RDS_MSG_RETRANSMITTED 5 355 #define RDS_MSG_MAPPED 6 356 #define RDS_MSG_PAGEVEC 7 357 #define RDS_MSG_FLUSH 8 358 359 struct rds_znotifier { 360 struct list_head z_list; 361 struct mmpin z_mmp; 362 u32 z_cookie; 363 }; 364 365 #define RDS_ZCOPY_SKB(__skb) ((struct rds_znotifier *)&((__skb)->cb[0])) 366 367 static inline struct sk_buff *rds_skb_from_znotifier(struct rds_znotifier *z) 368 { 369 return container_of((void *)z, struct sk_buff, cb); 370 } 371 372 struct rds_message { 373 refcount_t m_refcount; 374 struct list_head m_sock_item; 375 struct list_head m_conn_item; 376 struct rds_incoming m_inc; 377 u64 m_ack_seq; 378 __be32 m_daddr; 379 unsigned long m_flags; 380 381 /* Never access m_rs without holding m_rs_lock. 382 * Lock nesting is 383 * rm->m_rs_lock 384 * -> rs->rs_lock 385 */ 386 spinlock_t m_rs_lock; 387 wait_queue_head_t m_flush_wait; 388 389 struct rds_sock *m_rs; 390 391 /* cookie to send to remote, in rds header */ 392 rds_rdma_cookie_t m_rdma_cookie; 393 394 unsigned int m_used_sgs; 395 unsigned int m_total_sgs; 396 397 void *m_final_op; 398 399 struct { 400 struct rm_atomic_op { 401 int op_type; 402 union { 403 struct { 404 uint64_t compare; 405 uint64_t swap; 406 uint64_t compare_mask; 407 uint64_t swap_mask; 408 } op_m_cswp; 409 struct { 410 uint64_t add; 411 uint64_t nocarry_mask; 412 } op_m_fadd; 413 }; 414 415 u32 op_rkey; 416 u64 op_remote_addr; 417 unsigned int op_notify:1; 418 unsigned int op_recverr:1; 419 unsigned int op_mapped:1; 420 unsigned int op_silent:1; 421 unsigned int op_active:1; 422 struct scatterlist *op_sg; 423 struct rds_notifier *op_notifier; 424 425 struct rds_mr *op_rdma_mr; 426 } atomic; 427 struct rm_rdma_op { 428 u32 op_rkey; 429 u64 op_remote_addr; 430 unsigned int op_write:1; 431 unsigned int op_fence:1; 432 unsigned int op_notify:1; 433 unsigned int op_recverr:1; 434 unsigned int op_mapped:1; 435 unsigned int op_silent:1; 436 unsigned int op_active:1; 437 unsigned int op_bytes; 438 unsigned int op_nents; 439 unsigned int op_count; 440 struct scatterlist *op_sg; 441 struct rds_notifier *op_notifier; 442 443 struct rds_mr *op_rdma_mr; 444 } rdma; 445 struct rm_data_op { 446 unsigned int op_active:1; 447 unsigned int op_notify:1; 448 unsigned int op_nents; 449 unsigned int op_count; 450 unsigned int op_dmasg; 451 unsigned int op_dmaoff; 452 struct rds_znotifier *op_mmp_znotifier; 453 struct scatterlist *op_sg; 454 } data; 455 }; 456 }; 457 458 /* 459 * The RDS notifier is used (optionally) to tell the application about 460 * completed RDMA operations. Rather than keeping the whole rds message 461 * around on the queue, we allocate a small notifier that is put on the 462 * socket's notifier_list. Notifications are delivered to the application 463 * through control messages. 464 */ 465 struct rds_notifier { 466 struct list_head n_list; 467 uint64_t n_user_token; 468 int n_status; 469 }; 470 471 /** 472 * struct rds_transport - transport specific behavioural hooks 473 * 474 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send 475 * part of a message. The caller serializes on the send_sem so this 476 * doesn't need to be reentrant for a given conn. The header must be 477 * sent before the data payload. .xmit must be prepared to send a 478 * message with no data payload. .xmit should return the number of 479 * bytes that were sent down the connection, including header bytes. 480 * Returning 0 tells the caller that it doesn't need to perform any 481 * additional work now. This is usually the case when the transport has 482 * filled the sending queue for its connection and will handle 483 * triggering the rds thread to continue the send when space becomes 484 * available. Returning -EAGAIN tells the caller to retry the send 485 * immediately. Returning -ENOMEM tells the caller to retry the send at 486 * some point in the future. 487 * 488 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once 489 * it returns the connection can not call rds_recv_incoming(). 490 * This will only be called once after conn_connect returns 491 * non-zero success and will The caller serializes this with 492 * the send and connecting paths (xmit_* and conn_*). The 493 * transport is responsible for other serialization, including 494 * rds_recv_incoming(). This is called in process context but 495 * should try hard not to block. 496 */ 497 498 struct rds_transport { 499 char t_name[TRANSNAMSIZ]; 500 struct list_head t_item; 501 struct module *t_owner; 502 unsigned int t_prefer_loopback:1, 503 t_mp_capable:1; 504 unsigned int t_type; 505 506 int (*laddr_check)(struct net *net, __be32 addr); 507 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); 508 void (*conn_free)(void *data); 509 int (*conn_path_connect)(struct rds_conn_path *cp); 510 void (*conn_path_shutdown)(struct rds_conn_path *conn); 511 void (*xmit_path_prepare)(struct rds_conn_path *cp); 512 void (*xmit_path_complete)(struct rds_conn_path *cp); 513 int (*xmit)(struct rds_connection *conn, struct rds_message *rm, 514 unsigned int hdr_off, unsigned int sg, unsigned int off); 515 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); 516 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); 517 int (*recv_path)(struct rds_conn_path *cp); 518 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); 519 void (*inc_free)(struct rds_incoming *inc); 520 521 int (*cm_handle_connect)(struct rdma_cm_id *cm_id, 522 struct rdma_cm_event *event); 523 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); 524 void (*cm_connect_complete)(struct rds_connection *conn, 525 struct rdma_cm_event *event); 526 527 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, 528 unsigned int avail); 529 void (*exit)(void); 530 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, 531 struct rds_sock *rs, u32 *key_ret); 532 void (*sync_mr)(void *trans_private, int direction); 533 void (*free_mr)(void *trans_private, int invalidate); 534 void (*flush_mrs)(void); 535 bool (*t_unloading)(struct rds_connection *conn); 536 }; 537 538 struct rds_sock { 539 struct sock rs_sk; 540 541 u64 rs_user_addr; 542 u64 rs_user_bytes; 543 544 /* 545 * bound_addr used for both incoming and outgoing, no INADDR_ANY 546 * support. 547 */ 548 struct rhash_head rs_bound_node; 549 u64 rs_bound_key; 550 __be32 rs_bound_addr; 551 __be32 rs_conn_addr; 552 __be16 rs_bound_port; 553 __be16 rs_conn_port; 554 struct rds_transport *rs_transport; 555 556 /* 557 * rds_sendmsg caches the conn it used the last time around. 558 * This helps avoid costly lookups. 559 */ 560 struct rds_connection *rs_conn; 561 562 /* flag indicating we were congested or not */ 563 int rs_congested; 564 /* seen congestion (ENOBUFS) when sending? */ 565 int rs_seen_congestion; 566 567 /* rs_lock protects all these adjacent members before the newline */ 568 spinlock_t rs_lock; 569 struct list_head rs_send_queue; 570 u32 rs_snd_bytes; 571 int rs_rcv_bytes; 572 struct list_head rs_notify_queue; /* currently used for failed RDMAs */ 573 574 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask 575 * to decide whether the application should be woken up. 576 * If not set, we use rs_cong_track to find out whether a cong map 577 * update arrived. 578 */ 579 uint64_t rs_cong_mask; 580 uint64_t rs_cong_notify; 581 struct list_head rs_cong_list; 582 unsigned long rs_cong_track; 583 584 /* 585 * rs_recv_lock protects the receive queue, and is 586 * used to serialize with rds_release. 587 */ 588 rwlock_t rs_recv_lock; 589 struct list_head rs_recv_queue; 590 591 /* just for stats reporting */ 592 struct list_head rs_item; 593 594 /* these have their own lock */ 595 spinlock_t rs_rdma_lock; 596 struct rb_root rs_rdma_keys; 597 598 /* Socket options - in case there will be more */ 599 unsigned char rs_recverr, 600 rs_cong_monitor; 601 u32 rs_hash_initval; 602 603 /* Socket receive path trace points*/ 604 u8 rs_rx_traces; 605 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX]; 606 }; 607 608 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) 609 { 610 return container_of(sk, struct rds_sock, rs_sk); 611 } 612 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) 613 { 614 return &rs->rs_sk; 615 } 616 617 /* 618 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value 619 * to account for overhead. We don't account for overhead, we just apply 620 * the number of payload bytes to the specified value. 621 */ 622 static inline int rds_sk_sndbuf(struct rds_sock *rs) 623 { 624 return rds_rs_to_sk(rs)->sk_sndbuf / 2; 625 } 626 static inline int rds_sk_rcvbuf(struct rds_sock *rs) 627 { 628 return rds_rs_to_sk(rs)->sk_rcvbuf / 2; 629 } 630 631 struct rds_statistics { 632 uint64_t s_conn_reset; 633 uint64_t s_recv_drop_bad_checksum; 634 uint64_t s_recv_drop_old_seq; 635 uint64_t s_recv_drop_no_sock; 636 uint64_t s_recv_drop_dead_sock; 637 uint64_t s_recv_deliver_raced; 638 uint64_t s_recv_delivered; 639 uint64_t s_recv_queued; 640 uint64_t s_recv_immediate_retry; 641 uint64_t s_recv_delayed_retry; 642 uint64_t s_recv_ack_required; 643 uint64_t s_recv_rdma_bytes; 644 uint64_t s_recv_ping; 645 uint64_t s_send_queue_empty; 646 uint64_t s_send_queue_full; 647 uint64_t s_send_lock_contention; 648 uint64_t s_send_lock_queue_raced; 649 uint64_t s_send_immediate_retry; 650 uint64_t s_send_delayed_retry; 651 uint64_t s_send_drop_acked; 652 uint64_t s_send_ack_required; 653 uint64_t s_send_queued; 654 uint64_t s_send_rdma; 655 uint64_t s_send_rdma_bytes; 656 uint64_t s_send_pong; 657 uint64_t s_page_remainder_hit; 658 uint64_t s_page_remainder_miss; 659 uint64_t s_copy_to_user; 660 uint64_t s_copy_from_user; 661 uint64_t s_cong_update_queued; 662 uint64_t s_cong_update_received; 663 uint64_t s_cong_send_error; 664 uint64_t s_cong_send_blocked; 665 uint64_t s_recv_bytes_added_to_socket; 666 uint64_t s_recv_bytes_removed_from_socket; 667 668 }; 669 670 /* af_rds.c */ 671 void rds_sock_addref(struct rds_sock *rs); 672 void rds_sock_put(struct rds_sock *rs); 673 void rds_wake_sk_sleep(struct rds_sock *rs); 674 static inline void __rds_wake_sk_sleep(struct sock *sk) 675 { 676 wait_queue_head_t *waitq = sk_sleep(sk); 677 678 if (!sock_flag(sk, SOCK_DEAD) && waitq) 679 wake_up(waitq); 680 } 681 extern wait_queue_head_t rds_poll_waitq; 682 683 684 /* bind.c */ 685 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); 686 void rds_remove_bound(struct rds_sock *rs); 687 struct rds_sock *rds_find_bound(__be32 addr, __be16 port); 688 int rds_bind_lock_init(void); 689 void rds_bind_lock_destroy(void); 690 691 /* cong.c */ 692 int rds_cong_get_maps(struct rds_connection *conn); 693 void rds_cong_add_conn(struct rds_connection *conn); 694 void rds_cong_remove_conn(struct rds_connection *conn); 695 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); 696 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); 697 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); 698 void rds_cong_queue_updates(struct rds_cong_map *map); 699 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); 700 int rds_cong_updated_since(unsigned long *recent); 701 void rds_cong_add_socket(struct rds_sock *); 702 void rds_cong_remove_socket(struct rds_sock *); 703 void rds_cong_exit(void); 704 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); 705 706 /* conn.c */ 707 extern u32 rds_gen_num; 708 int rds_conn_init(void); 709 void rds_conn_exit(void); 710 struct rds_connection *rds_conn_create(struct net *net, 711 __be32 laddr, __be32 faddr, 712 struct rds_transport *trans, gfp_t gfp); 713 struct rds_connection *rds_conn_create_outgoing(struct net *net, 714 __be32 laddr, __be32 faddr, 715 struct rds_transport *trans, gfp_t gfp); 716 void rds_conn_shutdown(struct rds_conn_path *cpath); 717 void rds_conn_destroy(struct rds_connection *conn); 718 void rds_conn_drop(struct rds_connection *conn); 719 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy); 720 void rds_conn_connect_if_down(struct rds_connection *conn); 721 void rds_conn_path_connect_if_down(struct rds_conn_path *cp); 722 void rds_for_each_conn_info(struct socket *sock, unsigned int len, 723 struct rds_info_iterator *iter, 724 struct rds_info_lengths *lens, 725 int (*visitor)(struct rds_connection *, void *), 726 size_t item_len); 727 728 __printf(2, 3) 729 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...); 730 #define rds_conn_path_error(cp, fmt...) \ 731 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt) 732 733 static inline int 734 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new) 735 { 736 return atomic_cmpxchg(&cp->cp_state, old, new) == old; 737 } 738 739 static inline int 740 rds_conn_transition(struct rds_connection *conn, int old, int new) 741 { 742 WARN_ON(conn->c_trans->t_mp_capable); 743 return rds_conn_path_transition(&conn->c_path[0], old, new); 744 } 745 746 static inline int 747 rds_conn_path_state(struct rds_conn_path *cp) 748 { 749 return atomic_read(&cp->cp_state); 750 } 751 752 static inline int 753 rds_conn_state(struct rds_connection *conn) 754 { 755 WARN_ON(conn->c_trans->t_mp_capable); 756 return rds_conn_path_state(&conn->c_path[0]); 757 } 758 759 static inline int 760 rds_conn_path_up(struct rds_conn_path *cp) 761 { 762 return atomic_read(&cp->cp_state) == RDS_CONN_UP; 763 } 764 765 static inline int 766 rds_conn_up(struct rds_connection *conn) 767 { 768 WARN_ON(conn->c_trans->t_mp_capable); 769 return rds_conn_path_up(&conn->c_path[0]); 770 } 771 772 static inline int 773 rds_conn_path_connecting(struct rds_conn_path *cp) 774 { 775 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING; 776 } 777 778 static inline int 779 rds_conn_connecting(struct rds_connection *conn) 780 { 781 WARN_ON(conn->c_trans->t_mp_capable); 782 return rds_conn_path_connecting(&conn->c_path[0]); 783 } 784 785 /* message.c */ 786 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); 787 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); 788 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from, 789 bool zcopy); 790 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); 791 void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 792 __be16 dport, u64 seq); 793 int rds_message_add_extension(struct rds_header *hdr, 794 unsigned int type, const void *data, unsigned int len); 795 int rds_message_next_extension(struct rds_header *hdr, 796 unsigned int *pos, void *buf, unsigned int *buflen); 797 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); 798 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 799 void rds_message_inc_free(struct rds_incoming *inc); 800 void rds_message_addref(struct rds_message *rm); 801 void rds_message_put(struct rds_message *rm); 802 void rds_message_wait(struct rds_message *rm); 803 void rds_message_unmapped(struct rds_message *rm); 804 805 static inline void rds_message_make_checksum(struct rds_header *hdr) 806 { 807 hdr->h_csum = 0; 808 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); 809 } 810 811 static inline int rds_message_verify_checksum(const struct rds_header *hdr) 812 { 813 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; 814 } 815 816 817 /* page.c */ 818 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, 819 gfp_t gfp); 820 void rds_page_exit(void); 821 822 /* recv.c */ 823 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 824 __be32 saddr); 825 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn, 826 __be32 saddr); 827 void rds_inc_put(struct rds_incoming *inc); 828 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, 829 struct rds_incoming *inc, gfp_t gfp); 830 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, 831 int msg_flags); 832 void rds_clear_recv_queue(struct rds_sock *rs); 833 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); 834 void rds_inc_info_copy(struct rds_incoming *inc, 835 struct rds_info_iterator *iter, 836 __be32 saddr, __be32 daddr, int flip); 837 838 /* send.c */ 839 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); 840 void rds_send_path_reset(struct rds_conn_path *conn); 841 int rds_send_xmit(struct rds_conn_path *cp); 842 struct sockaddr_in; 843 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); 844 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); 845 void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 846 is_acked_func is_acked); 847 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, 848 is_acked_func is_acked); 849 void rds_send_ping(struct rds_connection *conn, int cp_index); 850 int rds_send_pong(struct rds_conn_path *cp, __be16 dport); 851 852 /* rdma.c */ 853 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); 854 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); 855 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); 856 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); 857 void rds_rdma_drop_keys(struct rds_sock *rs); 858 int rds_rdma_extra_size(struct rds_rdma_args *args); 859 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 860 struct cmsghdr *cmsg); 861 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, 862 struct cmsghdr *cmsg); 863 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 864 struct cmsghdr *cmsg); 865 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, 866 struct cmsghdr *cmsg); 867 void rds_rdma_free_op(struct rm_rdma_op *ro); 868 void rds_atomic_free_op(struct rm_atomic_op *ao); 869 void rds_rdma_send_complete(struct rds_message *rm, int wc_status); 870 void rds_atomic_send_complete(struct rds_message *rm, int wc_status); 871 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, 872 struct cmsghdr *cmsg); 873 874 void __rds_put_mr_final(struct rds_mr *mr); 875 static inline void rds_mr_put(struct rds_mr *mr) 876 { 877 if (refcount_dec_and_test(&mr->r_refcount)) 878 __rds_put_mr_final(mr); 879 } 880 881 static inline bool rds_destroy_pending(struct rds_connection *conn) 882 { 883 return !check_net(rds_conn_net(conn)) || 884 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn)); 885 } 886 887 /* stats.c */ 888 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); 889 #define rds_stats_inc_which(which, member) do { \ 890 per_cpu(which, get_cpu()).member++; \ 891 put_cpu(); \ 892 } while (0) 893 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) 894 #define rds_stats_add_which(which, member, count) do { \ 895 per_cpu(which, get_cpu()).member += count; \ 896 put_cpu(); \ 897 } while (0) 898 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) 899 int rds_stats_init(void); 900 void rds_stats_exit(void); 901 void rds_stats_info_copy(struct rds_info_iterator *iter, 902 uint64_t *values, const char *const *names, 903 size_t nr); 904 905 /* sysctl.c */ 906 int rds_sysctl_init(void); 907 void rds_sysctl_exit(void); 908 extern unsigned long rds_sysctl_sndbuf_min; 909 extern unsigned long rds_sysctl_sndbuf_default; 910 extern unsigned long rds_sysctl_sndbuf_max; 911 extern unsigned long rds_sysctl_reconnect_min_jiffies; 912 extern unsigned long rds_sysctl_reconnect_max_jiffies; 913 extern unsigned int rds_sysctl_max_unacked_packets; 914 extern unsigned int rds_sysctl_max_unacked_bytes; 915 extern unsigned int rds_sysctl_ping_enable; 916 extern unsigned long rds_sysctl_trace_flags; 917 extern unsigned int rds_sysctl_trace_level; 918 919 /* threads.c */ 920 int rds_threads_init(void); 921 void rds_threads_exit(void); 922 extern struct workqueue_struct *rds_wq; 923 void rds_queue_reconnect(struct rds_conn_path *cp); 924 void rds_connect_worker(struct work_struct *); 925 void rds_shutdown_worker(struct work_struct *); 926 void rds_send_worker(struct work_struct *); 927 void rds_recv_worker(struct work_struct *); 928 void rds_connect_path_complete(struct rds_conn_path *conn, int curr); 929 void rds_connect_complete(struct rds_connection *conn); 930 931 /* transport.c */ 932 void rds_trans_register(struct rds_transport *trans); 933 void rds_trans_unregister(struct rds_transport *trans); 934 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr); 935 void rds_trans_put(struct rds_transport *trans); 936 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, 937 unsigned int avail); 938 struct rds_transport *rds_trans_get(int t_type); 939 int rds_trans_init(void); 940 void rds_trans_exit(void); 941 942 #endif 943