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 mmpin z_mmp; 361 u32 z_cookie; 362 }; 363 364 struct rds_msg_zcopy_info { 365 struct list_head rs_zcookie_next; 366 union { 367 struct rds_znotifier znotif; 368 struct rds_zcopy_cookies zcookies; 369 }; 370 }; 371 372 struct rds_msg_zcopy_queue { 373 struct list_head zcookie_head; 374 spinlock_t lock; /* protects zcookie_head queue */ 375 }; 376 377 static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q) 378 { 379 spin_lock_init(&q->lock); 380 INIT_LIST_HEAD(&q->zcookie_head); 381 } 382 383 struct rds_message { 384 refcount_t m_refcount; 385 struct list_head m_sock_item; 386 struct list_head m_conn_item; 387 struct rds_incoming m_inc; 388 u64 m_ack_seq; 389 __be32 m_daddr; 390 unsigned long m_flags; 391 392 /* Never access m_rs without holding m_rs_lock. 393 * Lock nesting is 394 * rm->m_rs_lock 395 * -> rs->rs_lock 396 */ 397 spinlock_t m_rs_lock; 398 wait_queue_head_t m_flush_wait; 399 400 struct rds_sock *m_rs; 401 402 /* cookie to send to remote, in rds header */ 403 rds_rdma_cookie_t m_rdma_cookie; 404 405 unsigned int m_used_sgs; 406 unsigned int m_total_sgs; 407 408 void *m_final_op; 409 410 struct { 411 struct rm_atomic_op { 412 int op_type; 413 union { 414 struct { 415 uint64_t compare; 416 uint64_t swap; 417 uint64_t compare_mask; 418 uint64_t swap_mask; 419 } op_m_cswp; 420 struct { 421 uint64_t add; 422 uint64_t nocarry_mask; 423 } op_m_fadd; 424 }; 425 426 u32 op_rkey; 427 u64 op_remote_addr; 428 unsigned int op_notify:1; 429 unsigned int op_recverr:1; 430 unsigned int op_mapped:1; 431 unsigned int op_silent:1; 432 unsigned int op_active:1; 433 struct scatterlist *op_sg; 434 struct rds_notifier *op_notifier; 435 436 struct rds_mr *op_rdma_mr; 437 } atomic; 438 struct rm_rdma_op { 439 u32 op_rkey; 440 u64 op_remote_addr; 441 unsigned int op_write:1; 442 unsigned int op_fence:1; 443 unsigned int op_notify:1; 444 unsigned int op_recverr:1; 445 unsigned int op_mapped:1; 446 unsigned int op_silent:1; 447 unsigned int op_active:1; 448 unsigned int op_bytes; 449 unsigned int op_nents; 450 unsigned int op_count; 451 struct scatterlist *op_sg; 452 struct rds_notifier *op_notifier; 453 454 struct rds_mr *op_rdma_mr; 455 } rdma; 456 struct rm_data_op { 457 unsigned int op_active:1; 458 unsigned int op_notify:1; 459 unsigned int op_nents; 460 unsigned int op_count; 461 unsigned int op_dmasg; 462 unsigned int op_dmaoff; 463 struct rds_znotifier *op_mmp_znotifier; 464 struct scatterlist *op_sg; 465 } data; 466 }; 467 }; 468 469 /* 470 * The RDS notifier is used (optionally) to tell the application about 471 * completed RDMA operations. Rather than keeping the whole rds message 472 * around on the queue, we allocate a small notifier that is put on the 473 * socket's notifier_list. Notifications are delivered to the application 474 * through control messages. 475 */ 476 struct rds_notifier { 477 struct list_head n_list; 478 uint64_t n_user_token; 479 int n_status; 480 }; 481 482 /* Available as part of RDS core, so doesn't need to participate 483 * in get_preferred transport etc 484 */ 485 #define RDS_TRANS_LOOP 3 486 487 /** 488 * struct rds_transport - transport specific behavioural hooks 489 * 490 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send 491 * part of a message. The caller serializes on the send_sem so this 492 * doesn't need to be reentrant for a given conn. The header must be 493 * sent before the data payload. .xmit must be prepared to send a 494 * message with no data payload. .xmit should return the number of 495 * bytes that were sent down the connection, including header bytes. 496 * Returning 0 tells the caller that it doesn't need to perform any 497 * additional work now. This is usually the case when the transport has 498 * filled the sending queue for its connection and will handle 499 * triggering the rds thread to continue the send when space becomes 500 * available. Returning -EAGAIN tells the caller to retry the send 501 * immediately. Returning -ENOMEM tells the caller to retry the send at 502 * some point in the future. 503 * 504 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once 505 * it returns the connection can not call rds_recv_incoming(). 506 * This will only be called once after conn_connect returns 507 * non-zero success and will The caller serializes this with 508 * the send and connecting paths (xmit_* and conn_*). The 509 * transport is responsible for other serialization, including 510 * rds_recv_incoming(). This is called in process context but 511 * should try hard not to block. 512 */ 513 514 struct rds_transport { 515 char t_name[TRANSNAMSIZ]; 516 struct list_head t_item; 517 struct module *t_owner; 518 unsigned int t_prefer_loopback:1, 519 t_mp_capable:1; 520 unsigned int t_type; 521 522 int (*laddr_check)(struct net *net, __be32 addr); 523 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); 524 void (*conn_free)(void *data); 525 int (*conn_path_connect)(struct rds_conn_path *cp); 526 void (*conn_path_shutdown)(struct rds_conn_path *conn); 527 void (*xmit_path_prepare)(struct rds_conn_path *cp); 528 void (*xmit_path_complete)(struct rds_conn_path *cp); 529 int (*xmit)(struct rds_connection *conn, struct rds_message *rm, 530 unsigned int hdr_off, unsigned int sg, unsigned int off); 531 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); 532 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); 533 int (*recv_path)(struct rds_conn_path *cp); 534 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); 535 void (*inc_free)(struct rds_incoming *inc); 536 537 int (*cm_handle_connect)(struct rdma_cm_id *cm_id, 538 struct rdma_cm_event *event); 539 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id); 540 void (*cm_connect_complete)(struct rds_connection *conn, 541 struct rdma_cm_event *event); 542 543 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, 544 unsigned int avail); 545 void (*exit)(void); 546 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, 547 struct rds_sock *rs, u32 *key_ret); 548 void (*sync_mr)(void *trans_private, int direction); 549 void (*free_mr)(void *trans_private, int invalidate); 550 void (*flush_mrs)(void); 551 bool (*t_unloading)(struct rds_connection *conn); 552 }; 553 554 struct rds_sock { 555 struct sock rs_sk; 556 557 u64 rs_user_addr; 558 u64 rs_user_bytes; 559 560 /* 561 * bound_addr used for both incoming and outgoing, no INADDR_ANY 562 * support. 563 */ 564 struct rhash_head rs_bound_node; 565 u64 rs_bound_key; 566 __be32 rs_bound_addr; 567 __be32 rs_conn_addr; 568 __be16 rs_bound_port; 569 __be16 rs_conn_port; 570 struct rds_transport *rs_transport; 571 572 /* 573 * rds_sendmsg caches the conn it used the last time around. 574 * This helps avoid costly lookups. 575 */ 576 struct rds_connection *rs_conn; 577 578 /* flag indicating we were congested or not */ 579 int rs_congested; 580 /* seen congestion (ENOBUFS) when sending? */ 581 int rs_seen_congestion; 582 583 /* rs_lock protects all these adjacent members before the newline */ 584 spinlock_t rs_lock; 585 struct list_head rs_send_queue; 586 u32 rs_snd_bytes; 587 int rs_rcv_bytes; 588 struct list_head rs_notify_queue; /* currently used for failed RDMAs */ 589 590 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask 591 * to decide whether the application should be woken up. 592 * If not set, we use rs_cong_track to find out whether a cong map 593 * update arrived. 594 */ 595 uint64_t rs_cong_mask; 596 uint64_t rs_cong_notify; 597 struct list_head rs_cong_list; 598 unsigned long rs_cong_track; 599 600 /* 601 * rs_recv_lock protects the receive queue, and is 602 * used to serialize with rds_release. 603 */ 604 rwlock_t rs_recv_lock; 605 struct list_head rs_recv_queue; 606 607 /* just for stats reporting */ 608 struct list_head rs_item; 609 610 /* these have their own lock */ 611 spinlock_t rs_rdma_lock; 612 struct rb_root rs_rdma_keys; 613 614 /* Socket options - in case there will be more */ 615 unsigned char rs_recverr, 616 rs_cong_monitor; 617 u32 rs_hash_initval; 618 619 /* Socket receive path trace points*/ 620 u8 rs_rx_traces; 621 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX]; 622 struct rds_msg_zcopy_queue rs_zcookie_queue; 623 }; 624 625 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) 626 { 627 return container_of(sk, struct rds_sock, rs_sk); 628 } 629 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) 630 { 631 return &rs->rs_sk; 632 } 633 634 /* 635 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value 636 * to account for overhead. We don't account for overhead, we just apply 637 * the number of payload bytes to the specified value. 638 */ 639 static inline int rds_sk_sndbuf(struct rds_sock *rs) 640 { 641 return rds_rs_to_sk(rs)->sk_sndbuf / 2; 642 } 643 static inline int rds_sk_rcvbuf(struct rds_sock *rs) 644 { 645 return rds_rs_to_sk(rs)->sk_rcvbuf / 2; 646 } 647 648 struct rds_statistics { 649 uint64_t s_conn_reset; 650 uint64_t s_recv_drop_bad_checksum; 651 uint64_t s_recv_drop_old_seq; 652 uint64_t s_recv_drop_no_sock; 653 uint64_t s_recv_drop_dead_sock; 654 uint64_t s_recv_deliver_raced; 655 uint64_t s_recv_delivered; 656 uint64_t s_recv_queued; 657 uint64_t s_recv_immediate_retry; 658 uint64_t s_recv_delayed_retry; 659 uint64_t s_recv_ack_required; 660 uint64_t s_recv_rdma_bytes; 661 uint64_t s_recv_ping; 662 uint64_t s_send_queue_empty; 663 uint64_t s_send_queue_full; 664 uint64_t s_send_lock_contention; 665 uint64_t s_send_lock_queue_raced; 666 uint64_t s_send_immediate_retry; 667 uint64_t s_send_delayed_retry; 668 uint64_t s_send_drop_acked; 669 uint64_t s_send_ack_required; 670 uint64_t s_send_queued; 671 uint64_t s_send_rdma; 672 uint64_t s_send_rdma_bytes; 673 uint64_t s_send_pong; 674 uint64_t s_page_remainder_hit; 675 uint64_t s_page_remainder_miss; 676 uint64_t s_copy_to_user; 677 uint64_t s_copy_from_user; 678 uint64_t s_cong_update_queued; 679 uint64_t s_cong_update_received; 680 uint64_t s_cong_send_error; 681 uint64_t s_cong_send_blocked; 682 uint64_t s_recv_bytes_added_to_socket; 683 uint64_t s_recv_bytes_removed_from_socket; 684 685 }; 686 687 /* af_rds.c */ 688 void rds_sock_addref(struct rds_sock *rs); 689 void rds_sock_put(struct rds_sock *rs); 690 void rds_wake_sk_sleep(struct rds_sock *rs); 691 static inline void __rds_wake_sk_sleep(struct sock *sk) 692 { 693 wait_queue_head_t *waitq = sk_sleep(sk); 694 695 if (!sock_flag(sk, SOCK_DEAD) && waitq) 696 wake_up(waitq); 697 } 698 extern wait_queue_head_t rds_poll_waitq; 699 700 701 /* bind.c */ 702 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); 703 void rds_remove_bound(struct rds_sock *rs); 704 struct rds_sock *rds_find_bound(__be32 addr, __be16 port); 705 int rds_bind_lock_init(void); 706 void rds_bind_lock_destroy(void); 707 708 /* cong.c */ 709 int rds_cong_get_maps(struct rds_connection *conn); 710 void rds_cong_add_conn(struct rds_connection *conn); 711 void rds_cong_remove_conn(struct rds_connection *conn); 712 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); 713 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); 714 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); 715 void rds_cong_queue_updates(struct rds_cong_map *map); 716 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); 717 int rds_cong_updated_since(unsigned long *recent); 718 void rds_cong_add_socket(struct rds_sock *); 719 void rds_cong_remove_socket(struct rds_sock *); 720 void rds_cong_exit(void); 721 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); 722 723 /* conn.c */ 724 extern u32 rds_gen_num; 725 int rds_conn_init(void); 726 void rds_conn_exit(void); 727 struct rds_connection *rds_conn_create(struct net *net, 728 __be32 laddr, __be32 faddr, 729 struct rds_transport *trans, gfp_t gfp); 730 struct rds_connection *rds_conn_create_outgoing(struct net *net, 731 __be32 laddr, __be32 faddr, 732 struct rds_transport *trans, gfp_t gfp); 733 void rds_conn_shutdown(struct rds_conn_path *cpath); 734 void rds_conn_destroy(struct rds_connection *conn); 735 void rds_conn_drop(struct rds_connection *conn); 736 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy); 737 void rds_conn_connect_if_down(struct rds_connection *conn); 738 void rds_conn_path_connect_if_down(struct rds_conn_path *cp); 739 void rds_for_each_conn_info(struct socket *sock, unsigned int len, 740 struct rds_info_iterator *iter, 741 struct rds_info_lengths *lens, 742 int (*visitor)(struct rds_connection *, void *), 743 u64 *buffer, 744 size_t item_len); 745 746 __printf(2, 3) 747 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...); 748 #define rds_conn_path_error(cp, fmt...) \ 749 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt) 750 751 static inline int 752 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new) 753 { 754 return atomic_cmpxchg(&cp->cp_state, old, new) == old; 755 } 756 757 static inline int 758 rds_conn_transition(struct rds_connection *conn, int old, int new) 759 { 760 WARN_ON(conn->c_trans->t_mp_capable); 761 return rds_conn_path_transition(&conn->c_path[0], old, new); 762 } 763 764 static inline int 765 rds_conn_path_state(struct rds_conn_path *cp) 766 { 767 return atomic_read(&cp->cp_state); 768 } 769 770 static inline int 771 rds_conn_state(struct rds_connection *conn) 772 { 773 WARN_ON(conn->c_trans->t_mp_capable); 774 return rds_conn_path_state(&conn->c_path[0]); 775 } 776 777 static inline int 778 rds_conn_path_up(struct rds_conn_path *cp) 779 { 780 return atomic_read(&cp->cp_state) == RDS_CONN_UP; 781 } 782 783 static inline int 784 rds_conn_up(struct rds_connection *conn) 785 { 786 WARN_ON(conn->c_trans->t_mp_capable); 787 return rds_conn_path_up(&conn->c_path[0]); 788 } 789 790 static inline int 791 rds_conn_path_connecting(struct rds_conn_path *cp) 792 { 793 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING; 794 } 795 796 static inline int 797 rds_conn_connecting(struct rds_connection *conn) 798 { 799 WARN_ON(conn->c_trans->t_mp_capable); 800 return rds_conn_path_connecting(&conn->c_path[0]); 801 } 802 803 /* message.c */ 804 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); 805 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); 806 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from, 807 bool zcopy); 808 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); 809 void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 810 __be16 dport, u64 seq); 811 int rds_message_add_extension(struct rds_header *hdr, 812 unsigned int type, const void *data, unsigned int len); 813 int rds_message_next_extension(struct rds_header *hdr, 814 unsigned int *pos, void *buf, unsigned int *buflen); 815 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); 816 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 817 void rds_message_inc_free(struct rds_incoming *inc); 818 void rds_message_addref(struct rds_message *rm); 819 void rds_message_put(struct rds_message *rm); 820 void rds_message_wait(struct rds_message *rm); 821 void rds_message_unmapped(struct rds_message *rm); 822 void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info); 823 824 static inline void rds_message_make_checksum(struct rds_header *hdr) 825 { 826 hdr->h_csum = 0; 827 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); 828 } 829 830 static inline int rds_message_verify_checksum(const struct rds_header *hdr) 831 { 832 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; 833 } 834 835 836 /* page.c */ 837 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, 838 gfp_t gfp); 839 void rds_page_exit(void); 840 841 /* recv.c */ 842 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 843 __be32 saddr); 844 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn, 845 __be32 saddr); 846 void rds_inc_put(struct rds_incoming *inc); 847 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, 848 struct rds_incoming *inc, gfp_t gfp); 849 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, 850 int msg_flags); 851 void rds_clear_recv_queue(struct rds_sock *rs); 852 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); 853 void rds_inc_info_copy(struct rds_incoming *inc, 854 struct rds_info_iterator *iter, 855 __be32 saddr, __be32 daddr, int flip); 856 857 /* send.c */ 858 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); 859 void rds_send_path_reset(struct rds_conn_path *conn); 860 int rds_send_xmit(struct rds_conn_path *cp); 861 struct sockaddr_in; 862 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest); 863 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); 864 void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 865 is_acked_func is_acked); 866 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, 867 is_acked_func is_acked); 868 void rds_send_ping(struct rds_connection *conn, int cp_index); 869 int rds_send_pong(struct rds_conn_path *cp, __be16 dport); 870 871 /* rdma.c */ 872 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); 873 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); 874 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); 875 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); 876 void rds_rdma_drop_keys(struct rds_sock *rs); 877 int rds_rdma_extra_size(struct rds_rdma_args *args); 878 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 879 struct cmsghdr *cmsg); 880 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, 881 struct cmsghdr *cmsg); 882 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, 883 struct cmsghdr *cmsg); 884 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, 885 struct cmsghdr *cmsg); 886 void rds_rdma_free_op(struct rm_rdma_op *ro); 887 void rds_atomic_free_op(struct rm_atomic_op *ao); 888 void rds_rdma_send_complete(struct rds_message *rm, int wc_status); 889 void rds_atomic_send_complete(struct rds_message *rm, int wc_status); 890 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, 891 struct cmsghdr *cmsg); 892 893 void __rds_put_mr_final(struct rds_mr *mr); 894 static inline void rds_mr_put(struct rds_mr *mr) 895 { 896 if (refcount_dec_and_test(&mr->r_refcount)) 897 __rds_put_mr_final(mr); 898 } 899 900 static inline bool rds_destroy_pending(struct rds_connection *conn) 901 { 902 return !check_net(rds_conn_net(conn)) || 903 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn)); 904 } 905 906 /* stats.c */ 907 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); 908 #define rds_stats_inc_which(which, member) do { \ 909 per_cpu(which, get_cpu()).member++; \ 910 put_cpu(); \ 911 } while (0) 912 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) 913 #define rds_stats_add_which(which, member, count) do { \ 914 per_cpu(which, get_cpu()).member += count; \ 915 put_cpu(); \ 916 } while (0) 917 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) 918 int rds_stats_init(void); 919 void rds_stats_exit(void); 920 void rds_stats_info_copy(struct rds_info_iterator *iter, 921 uint64_t *values, const char *const *names, 922 size_t nr); 923 924 /* sysctl.c */ 925 int rds_sysctl_init(void); 926 void rds_sysctl_exit(void); 927 extern unsigned long rds_sysctl_sndbuf_min; 928 extern unsigned long rds_sysctl_sndbuf_default; 929 extern unsigned long rds_sysctl_sndbuf_max; 930 extern unsigned long rds_sysctl_reconnect_min_jiffies; 931 extern unsigned long rds_sysctl_reconnect_max_jiffies; 932 extern unsigned int rds_sysctl_max_unacked_packets; 933 extern unsigned int rds_sysctl_max_unacked_bytes; 934 extern unsigned int rds_sysctl_ping_enable; 935 extern unsigned long rds_sysctl_trace_flags; 936 extern unsigned int rds_sysctl_trace_level; 937 938 /* threads.c */ 939 int rds_threads_init(void); 940 void rds_threads_exit(void); 941 extern struct workqueue_struct *rds_wq; 942 void rds_queue_reconnect(struct rds_conn_path *cp); 943 void rds_connect_worker(struct work_struct *); 944 void rds_shutdown_worker(struct work_struct *); 945 void rds_send_worker(struct work_struct *); 946 void rds_recv_worker(struct work_struct *); 947 void rds_connect_path_complete(struct rds_conn_path *conn, int curr); 948 void rds_connect_complete(struct rds_connection *conn); 949 950 /* transport.c */ 951 void rds_trans_register(struct rds_transport *trans); 952 void rds_trans_unregister(struct rds_transport *trans); 953 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr); 954 void rds_trans_put(struct rds_transport *trans); 955 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, 956 unsigned int avail); 957 struct rds_transport *rds_trans_get(int t_type); 958 int rds_trans_init(void); 959 void rds_trans_exit(void); 960 961 #endif 962