1 /* SPDX-License-Identifier: GPL-2.0 */
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 #include <linux/in6.h>
18 * RDS Network protocol version
20 #define RDS_PROTOCOL_3_0 0x0300
21 #define RDS_PROTOCOL_3_1 0x0301
22 #define RDS_PROTOCOL_4_0 0x0400
23 #define RDS_PROTOCOL_4_1 0x0401
24 #define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
25 #define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
26 #define RDS_PROTOCOL_MINOR(v) ((v) & 255)
27 #define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
28 #define RDS_PROTOCOL_COMPAT_VERSION RDS_PROTOCOL_3_1
30 /* The following ports, 16385, 18634, 18635, are registered with IANA as
31 * the ports to be used for RDS over TCP and UDP. Currently, only RDS over
32 * TCP and RDS over IB/RDMA are implemented. 18634 is the historical value
33 * used for the RDMA_CM listener port. RDS/TCP uses port 16385. After
34 * IPv6 work, RDMA_CM also uses 16385 as the listener port. 18634 is kept
35 * to ensure compatibility with older RDS modules. Those ports are defined
36 * in each transport's header file.
38 #define RDS_PORT 18634
41 #define KERNEL_HAS_ATOMIC64
45 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
47 /* sigh, pr_debug() causes unused variable warnings */
48 static inline __printf(1, 2)
49 void rdsdebug(char *fmt, ...)
54 #define RDS_FRAG_SHIFT 12
55 #define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
57 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
58 #define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20))
60 #define RDS_CONG_MAP_BYTES (65536 / 8)
61 #define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
62 #define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
65 struct rb_node m_rb_node;
66 struct in6_addr m_addr;
67 wait_queue_head_t m_waitq;
68 struct list_head m_conn_list;
69 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
74 * This is how we will track the connection state:
75 * A connection is always in one of the following
76 * states. Updates to the state are atomic and imply
82 RDS_CONN_DISCONNECTING,
88 /* Bits for c_flags */
89 #define RDS_LL_SEND_FULL 0
90 #define RDS_RECONNECT_PENDING 1
92 #define RDS_RECV_REFILL 3
93 #define RDS_DESTROY_PENDING 4
95 /* Max number of multipaths per RDS connection. Must be a power of 2 */
96 #define RDS_MPATH_WORKERS 8
97 #define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
98 (rs)->rs_hash_initval) & ((n) - 1))
100 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
102 /* Per mpath connection state */
103 struct rds_conn_path {
104 struct rds_connection *cp_conn;
105 struct rds_message *cp_xmit_rm;
106 unsigned long cp_xmit_sg;
107 unsigned int cp_xmit_hdr_off;
108 unsigned int cp_xmit_data_off;
109 unsigned int cp_xmit_atomic_sent;
110 unsigned int cp_xmit_rdma_sent;
111 unsigned int cp_xmit_data_sent;
113 spinlock_t cp_lock; /* protect msg queues */
115 struct list_head cp_send_queue;
116 struct list_head cp_retrans;
120 void *cp_transport_data;
123 unsigned long cp_send_gen;
124 unsigned long cp_flags;
125 unsigned long cp_reconnect_jiffies;
126 struct delayed_work cp_send_w;
127 struct delayed_work cp_recv_w;
128 struct delayed_work cp_conn_w;
129 struct work_struct cp_down_w;
130 struct mutex cp_cm_lock; /* protect cp_state & cm */
131 wait_queue_head_t cp_waitq;
133 unsigned int cp_unacked_packets;
134 unsigned int cp_unacked_bytes;
135 unsigned int cp_index;
138 /* One rds_connection per RDS address pair */
139 struct rds_connection {
140 struct hlist_node c_hash_node;
141 struct in6_addr c_laddr;
142 struct in6_addr c_faddr;
143 int c_dev_if; /* ifindex used for this conn */
144 int c_bound_if; /* ifindex of c_laddr */
145 unsigned int c_loopback:1,
150 struct rds_connection *c_passive;
151 struct rds_transport *c_trans;
153 struct rds_cong_map *c_lcong;
154 struct rds_cong_map *c_fcong;
156 /* Protocol version */
157 unsigned int c_proposed_version;
158 unsigned int c_version;
159 possible_net_t c_net;
164 struct list_head c_map_item;
165 unsigned long c_map_queued;
167 struct rds_conn_path *c_path;
168 wait_queue_head_t c_hs_waitq; /* handshake waitq */
175 struct net *rds_conn_net(struct rds_connection *conn)
177 return read_pnet(&conn->c_net);
181 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
183 write_pnet(&conn->c_net, net);
186 #define RDS_FLAG_CONG_BITMAP 0x01
187 #define RDS_FLAG_ACK_REQUIRED 0x02
188 #define RDS_FLAG_RETRANSMITTED 0x04
189 #define RDS_MAX_ADV_CREDIT 255
191 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
192 * probe to exchange control information before establishing a connection.
193 * Currently the control information that is exchanged is the number of
194 * supported paths. If the peer is a legacy (older kernel revision) peer,
195 * it would return a pong message without additional control information
196 * that would then alert the sender that the peer was an older rev.
198 #define RDS_FLAG_PROBE_PORT 1
199 #define RDS_HS_PROBE(sport, dport) \
200 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
201 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
203 * Maximum space available for extension headers.
205 #define RDS_HEADER_EXT_SPACE 16
218 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
222 * Reserved - indicates end of extensions
224 #define RDS_EXTHDR_NONE 0
227 * This extension header is included in the very
228 * first message that is sent on a new connection,
229 * and identifies the protocol level. This will help
230 * rolling updates if a future change requires breaking
232 * NB: This is no longer true for IB, where we do a version
233 * negotiation during the connection setup phase (protocol
234 * version information is included in the RDMA CM private data).
236 #define RDS_EXTHDR_VERSION 1
237 struct rds_ext_header_version {
242 * This extension header is included in the RDS message
243 * chasing an RDMA operation.
245 #define RDS_EXTHDR_RDMA 2
246 struct rds_ext_header_rdma {
251 * This extension header tells the peer about the
252 * destination <R_Key,offset> of the requested RDMA
255 #define RDS_EXTHDR_RDMA_DEST 3
256 struct rds_ext_header_rdma_dest {
258 __be32 h_rdma_offset;
261 /* Extension header announcing number of paths.
262 * Implicit length = 2 bytes.
264 #define RDS_EXTHDR_NPATHS 5
265 #define RDS_EXTHDR_GEN_NUM 6
267 #define __RDS_EXTHDR_MAX 16 /* for now */
268 #define RDS_RX_MAX_TRACES (RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
269 #define RDS_MSG_RX_HDR 0
270 #define RDS_MSG_RX_START 1
271 #define RDS_MSG_RX_END 2
272 #define RDS_MSG_RX_CMSG 3
274 struct rds_incoming {
275 refcount_t i_refcount;
276 struct list_head i_item;
277 struct rds_connection *i_conn;
278 struct rds_conn_path *i_conn_path;
279 struct rds_header i_hdr;
280 unsigned long i_rx_jiffies;
281 struct in6_addr i_saddr;
283 rds_rdma_cookie_t i_rdma_cookie;
285 u64 i_rx_lat_trace[RDS_RX_MAX_TRACES];
289 struct rb_node r_rb_node;
290 refcount_t r_refcount;
293 /* A copy of the creation flags */
294 unsigned int r_use_once:1;
295 unsigned int r_invalidate:1;
296 unsigned int r_write:1;
298 /* This is for RDS_MR_DEAD.
299 * It would be nice & consistent to make this part of the above
300 * bit field here, but we need to use test_and_set_bit.
302 unsigned long r_state;
303 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
304 struct rds_transport *r_trans;
305 void *r_trans_private;
308 /* Flags for mr->r_state */
309 #define RDS_MR_DEAD 0
311 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
313 return r_key | (((u64) offset) << 32);
316 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
321 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
326 /* atomic operation types */
327 #define RDS_ATOMIC_TYPE_CSWP 0
328 #define RDS_ATOMIC_TYPE_FADD 1
331 * m_sock_item and m_conn_item are on lists that are serialized under
332 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
333 * the message will not be put back on the retransmit list after being sent.
334 * messages that are canceled while being sent rely on this.
336 * m_inc is used by loopback so that it can pass an incoming message straight
337 * back up into the rx path. It embeds a wire header which is also used by
338 * the send path, which is kind of awkward.
340 * m_sock_item indicates the message's presence on a socket's send or receive
341 * queue. m_rs will point to that socket.
343 * m_daddr is used by cancellation to prune messages to a given destination.
345 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
346 * nesting. As paths iterate over messages on a sock, or conn, they must
347 * also lock the conn, or sock, to remove the message from those lists too.
348 * Testing the flag to determine if the message is still on the lists lets
349 * us avoid testing the list_head directly. That means each path can use
350 * the message's list_head to keep it on a local list while juggling locks
351 * without confusing the other path.
353 * m_ack_seq is an optional field set by transports who need a different
354 * sequence number range to invalidate. They can use this in a callback
355 * that they pass to rds_send_drop_acked() to see if each message has been
356 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
357 * had ack_seq set yet.
359 #define RDS_MSG_ON_SOCK 1
360 #define RDS_MSG_ON_CONN 2
361 #define RDS_MSG_HAS_ACK_SEQ 3
362 #define RDS_MSG_ACK_REQUIRED 4
363 #define RDS_MSG_RETRANSMITTED 5
364 #define RDS_MSG_MAPPED 6
365 #define RDS_MSG_PAGEVEC 7
366 #define RDS_MSG_FLUSH 8
368 struct rds_znotifier {
373 struct rds_msg_zcopy_info {
374 struct list_head rs_zcookie_next;
376 struct rds_znotifier znotif;
377 struct rds_zcopy_cookies zcookies;
381 struct rds_msg_zcopy_queue {
382 struct list_head zcookie_head;
383 spinlock_t lock; /* protects zcookie_head queue */
386 static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
388 spin_lock_init(&q->lock);
389 INIT_LIST_HEAD(&q->zcookie_head);
392 struct rds_iov_vector {
393 struct rds_iovec *iov;
397 struct rds_iov_vector_arr {
398 struct rds_iov_vector *vec;
405 refcount_t m_refcount;
406 struct list_head m_sock_item;
407 struct list_head m_conn_item;
408 struct rds_incoming m_inc;
410 struct in6_addr m_daddr;
411 unsigned long m_flags;
413 /* Never access m_rs without holding m_rs_lock.
418 spinlock_t m_rs_lock;
419 wait_queue_head_t m_flush_wait;
421 struct rds_sock *m_rs;
423 /* cookie to send to remote, in rds header */
424 rds_rdma_cookie_t m_rdma_cookie;
426 unsigned int m_used_sgs;
427 unsigned int m_total_sgs;
432 struct rm_atomic_op {
438 uint64_t compare_mask;
443 uint64_t nocarry_mask;
449 unsigned int op_notify:1;
450 unsigned int op_recverr:1;
451 unsigned int op_mapped:1;
452 unsigned int op_silent:1;
453 unsigned int op_active:1;
454 struct scatterlist *op_sg;
455 struct rds_notifier *op_notifier;
457 struct rds_mr *op_rdma_mr;
462 unsigned int op_write:1;
463 unsigned int op_fence:1;
464 unsigned int op_notify:1;
465 unsigned int op_recverr:1;
466 unsigned int op_mapped:1;
467 unsigned int op_silent:1;
468 unsigned int op_active:1;
469 unsigned int op_bytes;
470 unsigned int op_nents;
471 unsigned int op_count;
472 struct scatterlist *op_sg;
473 struct rds_notifier *op_notifier;
475 struct rds_mr *op_rdma_mr;
478 unsigned int op_active:1;
479 unsigned int op_nents;
480 unsigned int op_count;
481 unsigned int op_dmasg;
482 unsigned int op_dmaoff;
483 struct rds_znotifier *op_mmp_znotifier;
484 struct scatterlist *op_sg;
488 struct rds_conn_path *m_conn_path;
492 * The RDS notifier is used (optionally) to tell the application about
493 * completed RDMA operations. Rather than keeping the whole rds message
494 * around on the queue, we allocate a small notifier that is put on the
495 * socket's notifier_list. Notifications are delivered to the application
496 * through control messages.
498 struct rds_notifier {
499 struct list_head n_list;
500 uint64_t n_user_token;
504 /* Available as part of RDS core, so doesn't need to participate
505 * in get_preferred transport etc
507 #define RDS_TRANS_LOOP 3
510 * struct rds_transport - transport specific behavioural hooks
512 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
513 * part of a message. The caller serializes on the send_sem so this
514 * doesn't need to be reentrant for a given conn. The header must be
515 * sent before the data payload. .xmit must be prepared to send a
516 * message with no data payload. .xmit should return the number of
517 * bytes that were sent down the connection, including header bytes.
518 * Returning 0 tells the caller that it doesn't need to perform any
519 * additional work now. This is usually the case when the transport has
520 * filled the sending queue for its connection and will handle
521 * triggering the rds thread to continue the send when space becomes
522 * available. Returning -EAGAIN tells the caller to retry the send
523 * immediately. Returning -ENOMEM tells the caller to retry the send at
524 * some point in the future.
526 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
527 * it returns the connection can not call rds_recv_incoming().
528 * This will only be called once after conn_connect returns
529 * non-zero success and will The caller serializes this with
530 * the send and connecting paths (xmit_* and conn_*). The
531 * transport is responsible for other serialization, including
532 * rds_recv_incoming(). This is called in process context but
533 * should try hard not to block.
536 struct rds_transport {
537 char t_name[TRANSNAMSIZ];
538 struct list_head t_item;
539 struct module *t_owner;
540 unsigned int t_prefer_loopback:1,
544 int (*laddr_check)(struct net *net, const struct in6_addr *addr,
546 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
547 void (*conn_free)(void *data);
548 int (*conn_path_connect)(struct rds_conn_path *cp);
549 void (*conn_path_shutdown)(struct rds_conn_path *conn);
550 void (*xmit_path_prepare)(struct rds_conn_path *cp);
551 void (*xmit_path_complete)(struct rds_conn_path *cp);
552 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
553 unsigned int hdr_off, unsigned int sg, unsigned int off);
554 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
555 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
556 int (*recv_path)(struct rds_conn_path *cp);
557 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
558 void (*inc_free)(struct rds_incoming *inc);
560 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
561 struct rdma_cm_event *event, bool isv6);
562 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
563 void (*cm_connect_complete)(struct rds_connection *conn,
564 struct rdma_cm_event *event);
566 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
569 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
570 struct rds_sock *rs, u32 *key_ret,
571 struct rds_connection *conn);
572 void (*sync_mr)(void *trans_private, int direction);
573 void (*free_mr)(void *trans_private, int invalidate);
574 void (*flush_mrs)(void);
575 bool (*t_unloading)(struct rds_connection *conn);
576 u8 (*get_tos_map)(u8 tos);
579 /* Bind hash table key length. It is the sum of the size of a struct
580 * in6_addr, a scope_id and a port.
582 #define RDS_BOUND_KEY_LEN \
583 (sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
592 * bound_addr used for both incoming and outgoing, no INADDR_ANY
595 struct rhash_head rs_bound_node;
596 u8 rs_bound_key[RDS_BOUND_KEY_LEN];
597 struct sockaddr_in6 rs_bound_sin6;
598 #define rs_bound_addr rs_bound_sin6.sin6_addr
599 #define rs_bound_addr_v4 rs_bound_sin6.sin6_addr.s6_addr32[3]
600 #define rs_bound_port rs_bound_sin6.sin6_port
601 #define rs_bound_scope_id rs_bound_sin6.sin6_scope_id
602 struct in6_addr rs_conn_addr;
603 #define rs_conn_addr_v4 rs_conn_addr.s6_addr32[3]
605 struct rds_transport *rs_transport;
608 * rds_sendmsg caches the conn it used the last time around.
609 * This helps avoid costly lookups.
611 struct rds_connection *rs_conn;
613 /* flag indicating we were congested or not */
615 /* seen congestion (ENOBUFS) when sending? */
616 int rs_seen_congestion;
618 /* rs_lock protects all these adjacent members before the newline */
620 struct list_head rs_send_queue;
623 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
625 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
626 * to decide whether the application should be woken up.
627 * If not set, we use rs_cong_track to find out whether a cong map
630 uint64_t rs_cong_mask;
631 uint64_t rs_cong_notify;
632 struct list_head rs_cong_list;
633 unsigned long rs_cong_track;
636 * rs_recv_lock protects the receive queue, and is
637 * used to serialize with rds_release.
639 rwlock_t rs_recv_lock;
640 struct list_head rs_recv_queue;
642 /* just for stats reporting */
643 struct list_head rs_item;
645 /* these have their own lock */
646 spinlock_t rs_rdma_lock;
647 struct rb_root rs_rdma_keys;
649 /* Socket options - in case there will be more */
650 unsigned char rs_recverr,
654 /* Socket receive path trace points*/
656 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
657 struct rds_msg_zcopy_queue rs_zcookie_queue;
661 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
663 return container_of(sk, struct rds_sock, rs_sk);
665 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
671 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
672 * to account for overhead. We don't account for overhead, we just apply
673 * the number of payload bytes to the specified value.
675 static inline int rds_sk_sndbuf(struct rds_sock *rs)
677 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
679 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
681 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
684 struct rds_statistics {
685 uint64_t s_conn_reset;
686 uint64_t s_recv_drop_bad_checksum;
687 uint64_t s_recv_drop_old_seq;
688 uint64_t s_recv_drop_no_sock;
689 uint64_t s_recv_drop_dead_sock;
690 uint64_t s_recv_deliver_raced;
691 uint64_t s_recv_delivered;
692 uint64_t s_recv_queued;
693 uint64_t s_recv_immediate_retry;
694 uint64_t s_recv_delayed_retry;
695 uint64_t s_recv_ack_required;
696 uint64_t s_recv_rdma_bytes;
697 uint64_t s_recv_ping;
698 uint64_t s_send_queue_empty;
699 uint64_t s_send_queue_full;
700 uint64_t s_send_lock_contention;
701 uint64_t s_send_lock_queue_raced;
702 uint64_t s_send_immediate_retry;
703 uint64_t s_send_delayed_retry;
704 uint64_t s_send_drop_acked;
705 uint64_t s_send_ack_required;
706 uint64_t s_send_queued;
707 uint64_t s_send_rdma;
708 uint64_t s_send_rdma_bytes;
709 uint64_t s_send_pong;
710 uint64_t s_page_remainder_hit;
711 uint64_t s_page_remainder_miss;
712 uint64_t s_copy_to_user;
713 uint64_t s_copy_from_user;
714 uint64_t s_cong_update_queued;
715 uint64_t s_cong_update_received;
716 uint64_t s_cong_send_error;
717 uint64_t s_cong_send_blocked;
718 uint64_t s_recv_bytes_added_to_socket;
719 uint64_t s_recv_bytes_removed_from_socket;
720 uint64_t s_send_stuck_rm;
724 void rds_sock_addref(struct rds_sock *rs);
725 void rds_sock_put(struct rds_sock *rs);
726 void rds_wake_sk_sleep(struct rds_sock *rs);
727 static inline void __rds_wake_sk_sleep(struct sock *sk)
729 wait_queue_head_t *waitq = sk_sleep(sk);
731 if (!sock_flag(sk, SOCK_DEAD) && waitq)
734 extern wait_queue_head_t rds_poll_waitq;
738 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
739 void rds_remove_bound(struct rds_sock *rs);
740 struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
742 int rds_bind_lock_init(void);
743 void rds_bind_lock_destroy(void);
746 int rds_cong_get_maps(struct rds_connection *conn);
747 void rds_cong_add_conn(struct rds_connection *conn);
748 void rds_cong_remove_conn(struct rds_connection *conn);
749 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
750 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
751 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
752 void rds_cong_queue_updates(struct rds_cong_map *map);
753 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
754 int rds_cong_updated_since(unsigned long *recent);
755 void rds_cong_add_socket(struct rds_sock *);
756 void rds_cong_remove_socket(struct rds_sock *);
757 void rds_cong_exit(void);
758 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
761 extern u32 rds_gen_num;
762 int rds_conn_init(void);
763 void rds_conn_exit(void);
764 struct rds_connection *rds_conn_create(struct net *net,
765 const struct in6_addr *laddr,
766 const struct in6_addr *faddr,
767 struct rds_transport *trans,
770 struct rds_connection *rds_conn_create_outgoing(struct net *net,
771 const struct in6_addr *laddr,
772 const struct in6_addr *faddr,
773 struct rds_transport *trans,
774 u8 tos, gfp_t gfp, int dev_if);
775 void rds_conn_shutdown(struct rds_conn_path *cpath);
776 void rds_conn_destroy(struct rds_connection *conn);
777 void rds_conn_drop(struct rds_connection *conn);
778 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
779 void rds_conn_connect_if_down(struct rds_connection *conn);
780 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
781 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
782 struct rds_info_iterator *iter,
783 struct rds_info_lengths *lens,
784 int (*visitor)(struct rds_connection *, void *),
789 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
790 #define rds_conn_path_error(cp, fmt...) \
791 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
794 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
796 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
800 rds_conn_transition(struct rds_connection *conn, int old, int new)
802 WARN_ON(conn->c_trans->t_mp_capable);
803 return rds_conn_path_transition(&conn->c_path[0], old, new);
807 rds_conn_path_state(struct rds_conn_path *cp)
809 return atomic_read(&cp->cp_state);
813 rds_conn_state(struct rds_connection *conn)
815 WARN_ON(conn->c_trans->t_mp_capable);
816 return rds_conn_path_state(&conn->c_path[0]);
820 rds_conn_path_up(struct rds_conn_path *cp)
822 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
826 rds_conn_up(struct rds_connection *conn)
828 WARN_ON(conn->c_trans->t_mp_capable);
829 return rds_conn_path_up(&conn->c_path[0]);
833 rds_conn_path_connecting(struct rds_conn_path *cp)
835 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
839 rds_conn_connecting(struct rds_connection *conn)
841 WARN_ON(conn->c_trans->t_mp_capable);
842 return rds_conn_path_connecting(&conn->c_path[0]);
846 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
847 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents,
849 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
851 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
852 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
853 __be16 dport, u64 seq);
854 int rds_message_add_extension(struct rds_header *hdr,
855 unsigned int type, const void *data, unsigned int len);
856 int rds_message_next_extension(struct rds_header *hdr,
857 unsigned int *pos, void *buf, unsigned int *buflen);
858 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
859 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
860 void rds_message_inc_free(struct rds_incoming *inc);
861 void rds_message_addref(struct rds_message *rm);
862 void rds_message_put(struct rds_message *rm);
863 void rds_message_wait(struct rds_message *rm);
864 void rds_message_unmapped(struct rds_message *rm);
865 void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
867 static inline void rds_message_make_checksum(struct rds_header *hdr)
870 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
873 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
875 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
880 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
882 void rds_page_exit(void);
885 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
886 struct in6_addr *saddr);
887 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
888 struct in6_addr *saddr);
889 void rds_inc_put(struct rds_incoming *inc);
890 void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
891 struct in6_addr *daddr,
892 struct rds_incoming *inc, gfp_t gfp);
893 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
895 void rds_clear_recv_queue(struct rds_sock *rs);
896 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
897 void rds_inc_info_copy(struct rds_incoming *inc,
898 struct rds_info_iterator *iter,
899 __be32 saddr, __be32 daddr, int flip);
900 void rds6_inc_info_copy(struct rds_incoming *inc,
901 struct rds_info_iterator *iter,
902 struct in6_addr *saddr, struct in6_addr *daddr,
906 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
907 void rds_send_path_reset(struct rds_conn_path *conn);
908 int rds_send_xmit(struct rds_conn_path *cp);
910 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest);
911 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
912 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
913 is_acked_func is_acked);
914 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
915 is_acked_func is_acked);
916 void rds_send_ping(struct rds_connection *conn, int cp_index);
917 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
920 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
921 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
922 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
923 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
924 void rds_rdma_drop_keys(struct rds_sock *rs);
925 int rds_rdma_extra_size(struct rds_rdma_args *args,
926 struct rds_iov_vector *iov);
927 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
928 struct cmsghdr *cmsg);
929 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
930 struct cmsghdr *cmsg,
931 struct rds_iov_vector *vec);
932 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
933 struct cmsghdr *cmsg);
934 void rds_rdma_free_op(struct rm_rdma_op *ro);
935 void rds_atomic_free_op(struct rm_atomic_op *ao);
936 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
937 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
938 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
939 struct cmsghdr *cmsg);
941 void __rds_put_mr_final(struct rds_mr *mr);
942 static inline void rds_mr_put(struct rds_mr *mr)
944 if (refcount_dec_and_test(&mr->r_refcount))
945 __rds_put_mr_final(mr);
948 static inline bool rds_destroy_pending(struct rds_connection *conn)
950 return !check_net(rds_conn_net(conn)) ||
951 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
955 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
956 #define rds_stats_inc_which(which, member) do { \
957 per_cpu(which, get_cpu()).member++; \
960 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
961 #define rds_stats_add_which(which, member, count) do { \
962 per_cpu(which, get_cpu()).member += count; \
965 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
966 int rds_stats_init(void);
967 void rds_stats_exit(void);
968 void rds_stats_info_copy(struct rds_info_iterator *iter,
969 uint64_t *values, const char *const *names,
973 int rds_sysctl_init(void);
974 void rds_sysctl_exit(void);
975 extern unsigned long rds_sysctl_sndbuf_min;
976 extern unsigned long rds_sysctl_sndbuf_default;
977 extern unsigned long rds_sysctl_sndbuf_max;
978 extern unsigned long rds_sysctl_reconnect_min_jiffies;
979 extern unsigned long rds_sysctl_reconnect_max_jiffies;
980 extern unsigned int rds_sysctl_max_unacked_packets;
981 extern unsigned int rds_sysctl_max_unacked_bytes;
982 extern unsigned int rds_sysctl_ping_enable;
983 extern unsigned long rds_sysctl_trace_flags;
984 extern unsigned int rds_sysctl_trace_level;
987 int rds_threads_init(void);
988 void rds_threads_exit(void);
989 extern struct workqueue_struct *rds_wq;
990 void rds_queue_reconnect(struct rds_conn_path *cp);
991 void rds_connect_worker(struct work_struct *);
992 void rds_shutdown_worker(struct work_struct *);
993 void rds_send_worker(struct work_struct *);
994 void rds_recv_worker(struct work_struct *);
995 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
996 void rds_connect_complete(struct rds_connection *conn);
997 int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
1000 void rds_trans_register(struct rds_transport *trans);
1001 void rds_trans_unregister(struct rds_transport *trans);
1002 struct rds_transport *rds_trans_get_preferred(struct net *net,
1003 const struct in6_addr *addr,
1005 void rds_trans_put(struct rds_transport *trans);
1006 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
1007 unsigned int avail);
1008 struct rds_transport *rds_trans_get(int t_type);
1009 int rds_trans_init(void);
1010 void rds_trans_exit(void);